candor_classify/policy.rs
1//! The canonical CANDOR_POLICY DSL parser (candor-spec SPEC §6.2).
2//!
3//! This is the **single** Rust implementation of the policy grammar — shared by the nightly dylint
4//! gate (`src/lib.rs`, AS-EFF-006/008/009) and the stable `candor-query` (`whatif`, and the
5//! `parsepolicy` dump the cross-impl conformance suite diffs against the JVM engine). Keeping one
6//! parser here is what makes "the gate means the same thing in every language" a fact rather than a
7//! hope: the Rust gate, the Rust pre-edit tool, and the cross-impl differential all read THIS code.
8//!
9//! Pure, stable Rust (string parsing only — no rustc types), so it lives beside the classifier.
10
11use crate::cap_from_name;
12use std::collections::{BTreeMap, BTreeSet};
13
14/// The honesty marker (SPEC §4). Denyable so `deny Unknown <scope>` forbids the *unverifiable* case.
15pub const UNKNOWN: &str = "Unknown";
16
17/// The NORMATIVE projection of a raw `unknown_why` reason onto a fixed, cross-engine reason CLASS
18/// (candor-spec REASON-SCOPED-UNKNOWN-DESIGN.md §1). Reason-scoped policies (`deny E Unknown[class]`)
19/// quantify over these classes, so the mapping MUST be identical in every engine — this mirrors the
20/// java reference `ReasonClass` (its `classify(String)` path, since rust emits raw string reasons). The
21/// class set is CLOSED (six members); a raw reason matching no pinned prefix maps to `Unresolved` —
22/// conservative: it stays in scope of any `Unknown[*]` / `Unknown[dynamic]` policy, never silently
23/// tolerated.
24#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
25pub enum ReasonClass {
26 /// reflection / metaprogramming
27 Reflect,
28 /// unresolved virtual/dynamic dispatch, same-name ambiguity, invokedynamic
29 Dispatch,
30 /// callback / closure / function-value / async-continuation indirection
31 Indirect,
32 /// FFI / native boundary
33 Native,
34 /// generic unresolvable call/import, AND the catch-all for any unrecognized raw reason
35 Unresolved,
36 /// analysis not wired up (fixable, not a real dynamic hole): missing-config / no-tsconfig
37 Setup,
38}
39
40impl ReasonClass {
41 /// The lowercase policy-facing token (`deny E Unknown[<token>]`).
42 pub fn token(self) -> &'static str {
43 match self {
44 ReasonClass::Reflect => "reflect",
45 ReasonClass::Dispatch => "dispatch",
46 ReasonClass::Indirect => "indirect",
47 ReasonClass::Native => "native",
48 ReasonClass::Unresolved => "unresolved",
49 ReasonClass::Setup => "setup",
50 }
51 }
52
53 /// Parse a policy-facing token back to a class; `None` if it names no class.
54 pub fn from_token(t: &str) -> Option<ReasonClass> {
55 Some(match t {
56 "reflect" => ReasonClass::Reflect,
57 "dispatch" => ReasonClass::Dispatch,
58 "indirect" => ReasonClass::Indirect,
59 "native" => ReasonClass::Native,
60 "unresolved" => ReasonClass::Unresolved,
61 "setup" => ReasonClass::Setup,
62 _ => return None,
63 })
64 }
65
66 /// Map a raw `unknown_why` reason to its normative class — prefix-based (raw reasons carry a
67 /// `kind:detail` shape, e.g. `dispatch:foo::Bar`), unrecognized → `Unresolved`. Byte-identical
68 /// intent to the java `ReasonClass.classify(String)`.
69 ///
70 /// ⟨0.24⟩ THIS IS THE ONLY PLACE THIS ENGINE HOLDS SPEC §4's KIND VOCABULARY. Every other reference
71 /// to a kind is either a raw string being emitted or the `dispatch:` prefix test in candor-query's
72 /// dispatch frontier; there is no typed kind enum here. §4's "AN ENGINE HOLDS THIS VOCABULARY TWICE,
73 /// AND THE HALVES DRIFT" paragraph records the JVM engine classifying `ambiguous` correctly HERE
74 /// while its typed `Kind` enum lacked the kind entirely — one token, two answers, inside one engine,
75 /// concealed precisely because this half was right. Holding it once is why that is not reachable in
76 /// this engine; a future typed representation must be added at the same commit as its control
77 /// (`off_vocabulary_kinds_round_trip_and_classify_through_the_catch_all`).
78 ///
79 /// The five §4 kinds are `reflect`/`native`/`dispatch`/`callback`/`ambiguous`. `ambiguous` maps to
80 /// `dispatch` and rust is its only PRODUCER; `indy`/`task-handoff` are candor-java's migration kinds
81 /// and `dep:`/`dep-stale:` are swift's registered per-dependency-ENTRY kinds, reaching `Unresolved`
82 /// through the catch-all, which is the class §6.2 prescribes for them.
83 pub fn classify(why: &str) -> ReasonClass {
84 let w = why.trim().to_ascii_lowercase();
85 if w.starts_with("reflect") || w == "dynamicmemberlookup" {
86 ReasonClass::Reflect
87 } else if w.starts_with("native") {
88 ReasonClass::Native
89 } else if w.starts_with("callback") || w.starts_with("closure") || w.starts_with("task-handoff") {
90 ReasonClass::Indirect
91 } else if w.starts_with("dispatch") || w.starts_with("indy") || w.starts_with("ambiguous") {
92 ReasonClass::Dispatch
93 } else if w.starts_with("missing-config") || w.starts_with("no-tsconfig") || w.starts_with("no-node_modules") {
94 ReasonClass::Setup
95 } else {
96 ReasonClass::Unresolved
97 }
98 }
99
100 /// The `dynamic` alias — every GENUINE blind-spot class (excludes `setup`), incl. `unresolved` (the
101 /// catch-all) so `Unknown[dynamic]` never under-gates. The design's recommended usable strict gate.
102 pub fn dynamic_set() -> BTreeSet<ReasonClass> {
103 [
104 ReasonClass::Reflect,
105 ReasonClass::Dispatch,
106 ReasonClass::Indirect,
107 ReasonClass::Native,
108 ReasonClass::Unresolved,
109 ]
110 .into_iter()
111 .collect()
112 }
113}
114
115/// SPEC §6.2 ⟨0.24⟩ — does a function's TRANSITIVE reason-class set intersect `want` (class tokens)?
116///
117/// THE `None`/EMPTY ARM IS THE FAIL-CLOSED NET, and it is why this lives here rather than being
118/// open-coded twice. §6.2: "a function whose `Unknown` carries no recorded reason CONTRIBUTES
119/// `unresolved` to its class set — so a narrowed filter never *silently* tolerates a hole it failed to
120/// classify." Read the other way round, `!contains ⇒ exclude` over an empty set drops the entry from
121/// EVERY filter including one naming its own class: a silent under-report wearing a filter.
122///
123/// `classes` is the ACCUMULATED (post-`propagate_str`) set, never the direct `unknownWhy`: a reason
124/// names a site in the function's OWN body (§4), so a function whose `Unknown` is purely inherited
125/// carries none, and matching against the direct field answers a different question. SHARED by the
126/// `deny E Unknown[class]` gate (candor-scan) and `unverified --class` (candor-query) so a gate and the
127/// disclosure explaining it cannot disagree about which holes a class names.
128///
129/// The empty arm is a NET, not a route: it keys on the ABSENCE of a class set, so any other reason on
130/// the same function hides whatever it was covering. The reasonless case that CAN co-occur with a
131/// reason — a direct `Unknown` the unit did not name — must therefore contribute `unresolved` into the
132/// DIRECT map before propagation (candor-scan's `unknown_via_dep`, candor-query's report-side signature),
133/// not arrive here by absence.
134pub fn reason_class_matches(classes: Option<&BTreeSet<String>>, want: &BTreeSet<&str>) -> bool {
135 match classes {
136 None => want.contains(ReasonClass::Unresolved.token()),
137 Some(cs) if cs.is_empty() => want.contains(ReasonClass::Unresolved.token()),
138 Some(cs) => cs.iter().any(|t| want.contains(t.as_str())),
139 }
140}
141
142/// Parse `unknown-alias <name> = <class,…>` lines from `.candor/config` text (⟨0.19⟩, SPEC §6.2) into a
143/// name→classes map. A name that shadows a built-in (`*`/`dynamic`/a class token) is warned-and-skipped (a
144/// config alias may not redefine a built-in), as is a definition naming no valid class. Byte-shape with the
145/// java reference `Config.addAlias`.
146pub fn parse_unknown_aliases(config_text: &str) -> std::collections::BTreeMap<String, BTreeSet<ReasonClass>> {
147 let mut out = std::collections::BTreeMap::new();
148 for raw in config_text.lines() {
149 let line = raw.split('#').next().unwrap_or("").trim();
150 if line.is_empty() {
151 continue;
152 }
153 let mut it = line.splitn(2, char::is_whitespace);
154 // Case-INSENSITIVE key match, like the config loaders in java/ts/swift (which lowercase the key) —
155 // a case-sensitive match here made `Unknown-Alias …` define an alias everywhere BUT rust (a four-way
156 // parse divergence; caught in review). The rest of the line (name + classes) stays case-sensitive.
157 if !it.next().is_some_and(|k| k.eq_ignore_ascii_case("unknown-alias")) {
158 continue;
159 }
160 let val = it.next().unwrap_or("").trim();
161 let Some((name, classes)) = val.split_once('=') else {
162 eprintln!("candor: ignoring `unknown-alias` (want `unknown-alias <name> = <class,…>`): {val}");
163 continue;
164 };
165 let name = name.trim();
166 if name.is_empty() || name == "*" || name == "dynamic" || ReasonClass::from_token(name).is_some() {
167 eprintln!("candor: ignoring `unknown-alias` with reserved/empty name `{name}` (may not shadow `*`/`dynamic`/a class token)");
168 continue;
169 }
170 let mut set = BTreeSet::new();
171 let mut bad: Vec<&str> = Vec::new();
172 for cn in classes.split(',') {
173 let cn = cn.trim();
174 if cn.is_empty() {
175 continue;
176 }
177 if cn == "dynamic" {
178 set.extend(ReasonClass::dynamic_set());
179 } else if let Some(rc) = ReasonClass::from_token(cn) {
180 set.insert(rc);
181 } else {
182 bad.push(cn);
183 }
184 }
185 // ⟨0.24⟩ **AN UNRECOGNISED TOKEN REFUSES THE WHOLE DEFINITION** — SPEC §6.2 `be0b9a9`: the rule
186 // binds every policy value list, and *"the second is the sharper one: the typo is in the
187 // vocabulary the policy is written against rather than in the policy itself, and it fails open
188 // identically."*
189 //
190 // MEASURED: `unknown-alias corp = dispatch,nativ` → the DEFINITION silently became `{dispatch}`,
191 // so `deny Unknown[corp]` exited 0 over a native-caused hole that `= dispatch,native` catches.
192 // The alias resolved, `used_aliases` recorded it, the verdict named the config — every disclosure
193 // fired correctly ABOUT A DEFINITION THAT WAS NOT THE ONE ON DISK.
194 //
195 // REFUSING THE DEFINITION rather than minting a new error channel is what makes this fit: an
196 // alias that does not exist is one the policy's own `Unknown[<name>]` cannot resolve, so it lands
197 // on the `errors` path already there and the gate routes refuse with exit 2 — naming the token
198 // AND the accepted set, as §6.2 requires. It also keeps the blast radius honest: a config
199 // defining ten aliases, one of them typo'd and NONE of them mentioned by this policy, changed
200 // nothing, and turning that into a red gate would be the mirror over-reach. The refusal is
201 // triggered by USE, exactly like `used_aliases` (recorded at the point of use, for the same
202 // reason).
203 if !bad.is_empty() {
204 eprintln!(
205 "candor: REFUSING `unknown-alias {name}` — it names unrecognised reason-class(es) `{}` \
206 (accepted: reflect, dispatch, indirect, native, unresolved, setup, plus `dynamic`). The \
207 definition is refused WHOLE rather than narrowed to the tokens that parsed: keeping the \
208 rest would silently redefine `{name}` as something narrower than the config says, and a \
209 policy using it would gate less than it reads. `{name}` is now undefined, so a policy \
210 naming it is a policy error (exit 2).",
211 bad.join("`, `")
212 );
213 } else if set.is_empty() {
214 eprintln!("candor: ignoring `unknown-alias {name}` — no valid reason-class");
215 } else {
216 out.insert(name.to_string(), set);
217 }
218 }
219 out
220}
221
222/// ⟨0.20⟩ Parse `net-partner <host>` lines from `.candor/config` (NET-DESTINATION-CLASS-DESIGN.md): the
223/// per-project set of business-partner hosts the `Net` destination-class classifier treats as
224/// `known-partner`. Multi-value (repeatable key); the value is host-normalized (`:port` stripped,
225/// lowercased) like `MODEL_HOSTS`. Case-insensitive key match, mirroring `parse_unknown_aliases` + the
226/// java/ts/swift config loaders. A partner is per-project — never a universal list.
227pub fn parse_net_partners(config_text: &str) -> BTreeSet<String> {
228 let mut out = BTreeSet::new();
229 for raw in config_text.lines() {
230 let line = raw.split('#').next().unwrap_or("").trim();
231 if line.is_empty() {
232 continue;
233 }
234 let mut it = line.splitn(2, char::is_whitespace);
235 if !it.next().is_some_and(|k| k.eq_ignore_ascii_case("net-partner")) {
236 continue;
237 }
238 let val = it.next().unwrap_or("").trim();
239 if !val.is_empty() {
240 out.insert(host_part(val).to_ascii_lowercase());
241 }
242 }
243 out
244}
245
246/// Discover `.candor/config` text for a policy/scan anchored at `start`: `$CANDOR_CONFIG` if set + readable,
247/// else the nearest `.candor/config` walking UP from `start`, else `None`. Read-only + lenient (no
248/// process-exit — the caller decides fail-closed); used to resolve `unknown-alias` for the §6.2 gate +
249/// `parsepolicy` so both reflect the same checked-in config.
250pub fn discover_config_text(start: &std::path::Path) -> Option<String> {
251 discover_config(start).map(|(_, t)| t)
252}
253
254/// ⟨0.24⟩ As [`discover_config_text`], but ALSO the PATH the text came from, canonicalized.
255///
256/// **WHY THE PATH IS NOW LOAD-BEARING** (SPEC §3.1): a `.candor/config` supplying `unknown-alias`
257/// vocabulary can move a verdict 0→1, and discovery walks PARENT DIRECTORIES, so a file anywhere above
258/// the policy participates — ambient, and until now invisible in the output. *"A verdict changed by a
259/// file the operator cannot see named in the output is the ambient-input failure this whole format
260/// exists to refuse; the remedy is the same one used everywhere else here — not to forbid the input, but
261/// to make it impossible for it to act unnamed."* So the gate document NAMES it, and the path has to
262/// travel out of discovery for that to be possible.
263///
264/// CANONICALIZED because the two routes reach the same file from different working directories, and
265/// §3.1's byte-equality MUST is about the DOCUMENT: a relative path would differ between them for no
266/// reason other than where each was invoked.
267pub fn discover_config(start: &std::path::Path) -> Option<(std::path::PathBuf, String)> {
268 let canon = |p: &std::path::Path| std::fs::canonicalize(p).unwrap_or_else(|_| p.to_path_buf());
269 if let Ok(p) = std::env::var("CANDOR_CONFIG") {
270 let p = std::path::PathBuf::from(p);
271 return std::fs::read_to_string(&p).ok().map(|t| (canon(&p), t));
272 }
273 let start = canon(start);
274 let mut cur = if start.is_dir() { Some(start.as_path()) } else { start.parent() };
275 while let Some(d) = cur {
276 let cand = d.join(".candor/config");
277 if cand.is_file() {
278 return std::fs::read_to_string(&cand).ok().map(|t| (canon(&cand), t));
279 }
280 cur = d.parent();
281 }
282 None
283}
284
285/// One `deny <Effect…> [scope]` / `pure <scope>` rule (AS-EFF-006). `effects` empty ⇒ a `pure` rule
286/// (ANY effect forbidden). `scope` is a path segment-scope the rule applies to (None = whole unit).
287#[derive(Debug, Clone)]
288pub struct PolicyRule {
289 pub effects: BTreeSet<&'static str>,
290 pub scope: Option<String>,
291 /// Reason-class filter on an `Unknown` membership (REASON-SCOPED-UNKNOWN-DESIGN.md): empty ⇒
292 /// `Unknown[*]` (any reason — the bare form); non-empty ⇒ the Unknown hit fires ONLY for a fn whose
293 /// (transitive) reason classes include one of these. Ignored when `effects` doesn't contain `Unknown`.
294 pub unknown_classes: BTreeSet<ReasonClass>,
295 /// Destination-class filter on a `Net` membership (NET-DESTINATION-CLASS-DESIGN.md): empty ⇒ `Net[*]`
296 /// (any destination — the bare form); non-empty ⇒ the Net hit fires ONLY for a fn whose (transitive)
297 /// destination classes include one of these. Ignored when `effects` doesn't contain `Net`.
298 pub net_classes: BTreeSet<String>,
299 pub raw: String,
300}
301
302/// One `allow <Effect> [in <scope>] <literal>…` rule (AS-EFF-008). The effect is one of the four
303/// that carry a literal surface (`Net` hosts / `Exec` commands / `Fs` paths / `Db` tables); a
304/// function in `scope` performing it may reach ONLY the listed literals. Matching is
305/// effect-specific (`literal_allowed`).
306#[derive(Debug, Clone)]
307pub struct AllowRule {
308 pub effect: &'static str,
309 pub scope: Option<String>,
310 pub literals: BTreeSet<String>,
311 pub raw: String,
312}
313
314/// One `forbid <A> -> <B>` module-layering rule (AS-EFF-009): a function in scope `A` must not
315/// transitively call into scope `B`.
316#[derive(Debug, Clone)]
317pub struct LayerRule {
318 pub from: String,
319 pub to: String,
320 pub raw: String,
321}
322
323/// ⟨0.24⟩ ONE POLICY LINE THE PARSER DID NOT HONOUR AS WRITTEN (SPEC §3.1 `901f14d` / `195d45a`).
324///
325/// **THE DEFECT.** `parsepolicy` emitted **no `errors` key at all** (measured 2026-07-28 on the
326/// conformance battery: java 10, ts 4, rust 0, swift 0). Every one of these facts existed — they were
327/// printed to stderr as "ignoring policy rule …" — so the verb whose entire purpose is to let a consumer
328/// diff what an engine made of a policy answered the question with the not-honoured half deleted. Worse,
329/// it was INCONSISTENT with this engine's own gate: the gate refuses an unrecognised class token while
330/// the parse narrowed it silently, which is two answers to one question.
331///
332/// **`kind` IS A CLOSED SET, AND IT IS THE SPEC'S, NOT THE REFERENCE ENGINE'S.** `901f14d` pins four
333/// values: `reason-class/alias`, `Net destination-class`, `effect-name`, `rule-kind`. Measured, candor-java
334/// emits `forbid form`, `allow values` and `rule kind` (space, not hyphen) — three values outside the set
335/// and one spelling divergence — and candor-ts renames `kind`→`vocabulary` and `rule`→`where` and emits
336/// `accepted` as a PROSE STRING. This engine follows the clause: a line that names a rule keyword but does
337/// not form that keyword's rule formed no rule kind, so it is `rule-kind`.
338///
339/// `accepted` is an ARRAY OF TOKENS — the tokens that WOULD have been honoured in the position the bad one
340/// occupies. Empty where the position is open-ended (a host, a path), which is a fact about the grammar
341/// rather than a gap in the report.
342#[derive(Debug, Clone)]
343pub struct PolicyError {
344 /// One of [`PolicyError::KIND_REASON_CLASS`], [`PolicyError::KIND_NET_CLASS`],
345 /// [`PolicyError::KIND_EFFECT_NAME`], [`PolicyError::KIND_RULE_KIND`].
346 pub kind: &'static str,
347 /// The offending token, verbatim. Empty when the position was EMPTY (a missing arrow, an `allow` with
348 /// no values) — which is itself the finding.
349 pub token: String,
350 /// The tokens accepted in that position. Empty ⇒ the position takes an open-ended literal.
351 pub accepted: Vec<String>,
352 /// The raw policy line, verbatim.
353 pub rule: String,
354 /// The human sentence — the same text the stderr channel carries, so the two cannot disagree.
355 pub message: String,
356 /// ⟨0.24⟩ Does this error make the policy UNHONOURABLE, so every gate route must refuse (exit 2)?
357 ///
358 /// FATAL and REPORTED are different questions and this field is the only place they are told apart.
359 /// A dropped `nonsense line` is reported and survivable — the rest of the policy means what it says.
360 /// A rewritten `deny Unknown[dispatch,nativ]` is not: the rule that RAN is not the rule that was
361 /// written, and the direction that matters NARROWS it.
362 pub fatal: bool,
363}
364
365impl PolicyError {
366 pub const KIND_REASON_CLASS: &'static str = "reason-class/alias";
367 pub const KIND_NET_CLASS: &'static str = "Net destination-class";
368 pub const KIND_EFFECT_NAME: &'static str = "effect-name";
369 pub const KIND_RULE_KIND: &'static str = "rule-kind";
370}
371
372/// The rule kinds parsed from a CANDOR_POLICY file.
373#[derive(Default, Debug)]
374pub struct ParsedPolicy {
375 pub rules: Vec<PolicyRule>,
376 pub allow_rules: Vec<AllowRule>,
377 pub layer_rules: Vec<LayerRule>,
378 /// ⟨0.24⟩ POLICY ERRORS — a policy that cannot be honoured AS WRITTEN (SPEC §6.2). Non-empty ⇒ every
379 /// gate route MUST refuse: exit 2, the unreadable-policy posture. Not a warning list: the rules in
380 /// `rules` are what the text would mean if the error were tolerated, and tolerating it is the defect.
381 ///
382 /// Today the only member is an unrecognised reason-class/alias token in an `Unknown[…]` filter. The
383 /// asymmetry that used to justify a warning — "a dropped policy token leaves a WIDER rule standing,
384 /// so the failure is loud" — is false in the case that matters, and the false half is FAIL-OPEN:
385 ///
386 /// - `deny Unknown[corp]` (sole unrecognised token) — the filter empties and the rule WIDENS to a
387 /// bare `deny Unknown`, while the engine prints "ignoring policy rule" and then KEEPS and
388 /// re-scopes it. Merely surprising, but a FALSE DISCLOSURE.
389 /// - `deny Unknown[dispatch,nativ]` (a typo BESIDE valid tokens) — the token is dropped, the rule
390 /// NARROWS to `[dispatch]`, and it stops gating native-caused holes entirely while the operator
391 /// reads a gate that looks armed. **That is the fail-open, and it is the common case: a typo
392 /// lands beside correct tokens far more often than alone.**
393 ///
394 /// A policy that cannot be honoured as written is not silently rewritten into a different policy.
395 ///
396 /// ⟨0.24⟩ THIS LIST NOW HOLDS EVERY LINE THE PARSER DID NOT HONOUR, fatal or not (SPEC §3.1
397 /// `195d45a`) — `parsepolicy` reports them all, and the gate routes refuse on
398 /// [`ParsedPolicy::fatal_messages`] alone. Widening the LIST without widening what REFUSES is the
399 /// whole of the change: a dropped `nonsense line` was always survivable and stays so.
400 pub errors: Vec<PolicyError>,
401 /// ⟨0.24⟩ The `.candor/config` `unknown-alias` definitions this policy actually resolved a token
402 /// through (SPEC §3.1) — **name → the reason-class TOKENS it expanded to**, not a bare name list.
403 /// Non-empty ⇒ a config file supplied vocabulary that PARTICIPATED in the verdict, and the
404 /// `--gate-json` document MUST name that file. Recorded at the point of USE, not from the alias map:
405 /// a config defining ten aliases none of which the policy mentions changed nothing, and naming it
406 /// would train the reader to ignore the field.
407 ///
408 /// ⟨0.24⟩ **THE VALUE TRAVELS WITH THE NAME, AND THAT IS A SPEC MUST** (§3.1, candor-spec `7f5b5ba`).
409 /// This engine shipped the bare name — as did java and swift — and candor-ts kept the map and won the
410 /// argument from the clause's OWN sentence: `configSources: [path]` is rejected there because *a
411 /// disclosure that names the source but not the content leaves the reader knowing they were affected
412 /// and not how*, and `["corp"]` fails that same test one level down. **`corp = reflect` and
413 /// `corp = reflect,native` gate DIFFERENTLY under one unchanged policy line**, so a reader given only
414 /// the name cannot tell which gate ran. The map is a strict superset — the keys recover the old array.
415 ///
416 /// Class TOKENS rather than `ReasonClass`, so the wire order is the token's alphabetical one (which
417 /// is what candor-ts's `[...set].sort()` produces) and not `ReasonClass`'s declaration order.
418 pub used_aliases: BTreeMap<String, BTreeSet<String>>,
419}
420
421impl ParsedPolicy {
422 /// ⟨0.24⟩ The messages of the errors that make the policy UNHONOURABLE — what every gate route
423 /// refuses on. Non-empty ⇒ exit 2, the unreadable-policy posture.
424 ///
425 /// Separate from `errors` because REPORTED and FATAL are different questions, and conflating them in
426 /// either direction is a defect: refusing on every dropped line would make `nonsense line` fail a
427 /// build, and reporting only the fatal ones is the silent narrowing this rung exists to close.
428 pub fn fatal_messages(&self) -> Vec<&str> {
429 self.errors.iter().filter(|e| e.fatal).map(|e| e.message.as_str()).collect()
430 }
431}
432
433/// The hostname part of a `host[:port]` literal, port stripped — so `api.stripe.com` in a rule accepts
434/// a reached `api.stripe.com:443`. IPv6-aware: a bracketed `[host]:port` yields the bracketed host, and
435/// a BARE IPv6 literal (>1 colon, no brackets) has no port to strip and is returned whole — a naive
436/// first-colon split collapsed every `2001:db8::*` to `2001`, so one allowed IPv6 accepted any address
437/// in that block (/code-review). A hostname/IPv4 `host` or `host:port` (≤1 colon) splits at the colon.
438pub fn host_part(h: &str) -> &str {
439 if let Some(rest) = h.strip_prefix('[') {
440 // `[ipv6]` or `[ipv6]:port` — the host is between the brackets.
441 return rest.split(']').next().unwrap_or(rest);
442 }
443 if h.matches(':').count() > 1 {
444 return h; // bare IPv6 literal — no port suffix to strip
445 }
446 h.split(':').next().unwrap_or(h)
447}
448
449/// The basename of a command (`/usr/bin/git` → `git`), so `allow Exec … git` accepts an absolute path.
450pub fn cmd_base(c: &str) -> &str {
451 c.rsplit(['/', '\\']).next().unwrap_or(c)
452}
453
454/// Whether an allowed path `a` covers a reached path `r` (SPEC §6.2: path-boundary-respecting prefix).
455/// A directory covers itself and everything beneath it, but NOT a sibling sharing a textual prefix
456/// (`/etc/app` ⊉ `/etc/apppwned`); a `..` that climbs out is never covered; absolute/relative are
457/// never conflated.
458pub fn fs_path_covered(a: &str, r: &str) -> bool {
459 if r.split(['/', '\\']).any(|c| c == "..") {
460 return false;
461 }
462 let absolute = |s: &str| s.starts_with('/') || s.starts_with('\\');
463 if absolute(a) != absolute(r) {
464 return false;
465 }
466 let norm = |s: &str| -> Vec<String> {
467 s.split(['/', '\\'])
468 .filter(|c| !c.is_empty() && *c != ".")
469 .map(|c| c.to_string())
470 .collect()
471 };
472 let (ac, rc) = (norm(a), norm(r));
473 ac.len() <= rc.len() && ac.iter().zip(&rc).all(|(x, y)| x == y)
474}
475
476/// Whether an allowed table entry `a` covers a reached table `r` (SPEC §6.2): case-insensitive
477/// exact match on the (possibly schema-qualified) name, or a `schema.*` entry covering every table
478/// in that schema. Strict on qualification — an allowed `entries` does NOT cover a reached
479/// `ledger.entries` (write both forms if your queries mix them); silent widening is the failure
480/// mode an allowlist exists to prevent.
481pub fn db_table_covered(a: &str, r: &str) -> bool {
482 let (a, r) = (a.to_lowercase(), r.to_lowercase());
483 if let Some(schema) = a.strip_suffix(".*") {
484 return r.strip_prefix(schema).is_some_and(|rest| rest.starts_with('.'));
485 }
486 a == r
487}
488
489/// Whether a reached literal is allowed under an effect-specific match (SPEC §6.2): `Net` host by
490/// name (port ignored), `Exec` command by basename, `Fs` path by boundary-respecting prefix,
491/// `Db` table by qualified name or `schema.*`.
492pub fn literal_allowed(effect: &str, reached: &str, allow: &BTreeSet<String>) -> bool {
493 match effect {
494 // `Llm` ⟨0.13⟩ rides the Net host literal (SPEC §1) — matched by hostname like `Net`.
495 "Net" | "Llm" => allow.iter().any(|a| host_part(a) == host_part(reached)),
496 "Exec" => allow.iter().any(|a| cmd_base(a) == cmd_base(reached)),
497 "Fs" => allow.iter().any(|a| fs_path_covered(a, reached)),
498 "Db" => allow.iter().any(|a| db_table_covered(a, reached)),
499 _ => allow.contains(reached),
500 }
501}
502
503/// Split a function name (or scope) into PATH SEGMENTS on either separator. Reports reach the Rust gate
504/// AND `candor-query` from BOTH the Rust engines (`::`-separated names) and the JVM/Swift/TS engines
505/// (`.`-separated names — `candor-query` is explicitly built to read them). Segmenting on `::` ALONE
506/// left a scoped `deny`/`pure` rule silently INERT on a dotted name: the scope matched nothing, so
507/// `whatif` returned a false green on the security boundary (gate-evasion). The JVM engine's own
508/// `scopeMatches` already splits on `.`; this aligns the Rust side. A `:`/`.` never appears WITHIN a
509/// real segment, so splitting on both never over-segments a Rust name (no spurious match).
510fn name_segments(s: &str) -> Vec<&str> {
511 s.split(['.', ':']).filter(|p| !p.is_empty()).collect()
512}
513
514/// A policy scope matches a function name by **path segment** (SPEC §6.2), not substring: split both
515/// into segments (on `::` or `.`); the scope matches a contiguous run of name-segments where every
516/// segment except the last matches exactly and the last is a prefix. So `domain` matches
517/// `app::domain::h`, `com.acme.domain.h`, and `domain_logic` but not `subdomain`.
518pub fn scope_matches(name: &str, scope: &str) -> bool {
519 let segs = name_segments(name);
520 let parts = name_segments(scope);
521 if parts.is_empty() || parts.len() > segs.len() {
522 return false;
523 }
524 let (last, init) = parts.split_last().unwrap();
525 segs.windows(parts.len()).any(|w| {
526 let (w_last, w_init) = w.split_last().unwrap();
527 w_init == init && w_last.starts_with(last)
528 })
529}
530
531/// Reconstruct a rule's source form and the `Unknown`-forbidding upgrade for it: `pure <scope>` →
532/// (`"pure <scope>"`, `"deny Unknown <scope>"`); `deny <E…> <scope>` → (`"deny <E…> <scope>"`,
533/// `"deny <E…> Unknown <scope>"`). Shared so the gate note and `candor unverified` name the identical
534/// rule and upgrade — one source of truth for the disclosure's advice.
535///
536/// ⟨0.24⟩ **THE NARROWING FILTERS ARE RENDERED, and they had to start being rendered in the same commit
537/// that made them REACHABLE here.** Making `unverified_hole_rule` filter-aware is what first lets a
538/// `deny Unknown[reflect]` / `deny Net[unknown-host]` rule be the rule a hole is disclosed under — and
539/// this reconstruction dropped the bracket, so the fix would have printed the operator's narrowed rule
540/// back to them as the WIDE one (`deny Unknown`) and advised the nonsense upgrade `deny Unknown
541/// Unknown`. That is the same mis-attribution `whatif` carries, arriving through the fix for a different
542/// defect: the hazard on this rung is that each correction manufactures its own mirror, so the rendering
543/// moves with the predicate rather than after it.
544///
545/// A rule carrying NO filter renders byte-identically to before, which is what keeps conformance PARTs
546/// 12c/12d (`deny Db Net Unknown domain`, four-way) unmoved.
547///
548/// THE UPGRADE SPLITS on whether the rule already denies `Unknown`. If it does, it can only be here
549/// NARROWED — a bare `deny … Unknown` fires on every `Unknown`, so the function would be a violation and
550/// not a hole — and the upgrade is that term WIDENED to bare `Unknown`, not a second `Unknown` appended.
551pub fn rule_and_upgrade(r: &PolicyRule) -> (String, String) {
552 let scope = r.scope.clone().unwrap_or_default();
553 let suffix = if scope.is_empty() { String::new() } else { format!(" {scope}") };
554 if r.effects.is_empty() {
555 // `pure` forbids real effects but not Unknown; to REQUIRE provable purity, add a deny-Unknown.
556 return (format!("pure{suffix}"), format!("deny Unknown{suffix}"));
557 }
558 // One effect term, with its narrowing filter if it has one. Class tokens sorted by TOKEN string, as
559 // `parsepolicy` sorts them and as the java reference's `.sorted()` does — the dump and the
560 // disclosure must spell one rule one way.
561 let term = |e: &str| -> String {
562 if e == UNKNOWN && !r.unknown_classes.is_empty() {
563 let mut t: Vec<&str> = r.unknown_classes.iter().map(|c| c.token()).collect();
564 t.sort_unstable();
565 format!("{UNKNOWN}[{}]", t.join(","))
566 } else if e == "Net" && !r.net_classes.is_empty() {
567 format!("Net[{}]", r.net_classes.iter().map(String::as_str).collect::<Vec<_>>().join(","))
568 } else {
569 e.to_string()
570 }
571 };
572 let effs = r.effects.iter().map(|e| term(e)).collect::<Vec<_>>().join(" ");
573 if r.effects.contains(UNKNOWN) {
574 let widened =
575 r.effects.iter().map(|e| if *e == UNKNOWN { UNKNOWN.to_string() } else { term(e) }).collect::<Vec<_>>();
576 (format!("deny {effs}{suffix}"), format!("deny {}{suffix}", widened.join(" ")))
577 } else {
578 (format!("deny {effs}{suffix}"), format!("deny {effs} {UNKNOWN}{suffix}"))
579 }
580}
581
582/// The single predicate for a provable-purity hole (eval/fixloop/DISPATCH-NOTE.md): a function that is
583/// `Unknown`, sits in a `pure`/`deny <E>` scope, and PASSES that rule (carries none of its forbidden real
584/// effects) — so its compliance is asserted but not verified (the Unknown could hide the very effect the
585/// rule forbids; the classic case is a fn/closure-injected port). A *real* violation is the gate's job, not
586/// this. Returns the first governing rule under which the function is such a hole, or `None` if it is not
587/// one. Shared by candor-scan's gate note and candor-query's `unverified` so "what a hole is" has ONE
588/// definition — the two paths can never drift (conformance PART 12d pins their agreement).
589///
590/// ⟨0.24⟩ **"PASSES" IS NOW ASKED OF THE GATE, NOT OF A SECOND COPY OF IT.** This predicate computed the
591/// passing test from `r.effects` alone — "does the rule NAME an effect this function has?" — which is
592/// the pre-⟨0.19⟩ question, still being asked after two rungs gave rules a NARROWING FILTER. So on
593/// `deny Unknown[reflect]` over an `indirect` hole the gate TOLERATED (exit 0, the class does not match)
594/// while this read the same rule as violated; and a hole is *by definition* a function that PASSES its
595/// rule while `Unknown`, so the real hole was reclassified as a violation-that-isn't and **deleted from
596/// the disclosure** — `unverified` answering "every function in a pure/deny layer is PROVABLY clean ✓"
597/// over a function the gate had just declined to clear. MEASURED 2026-07-28, and reachable with no alias
598/// in play at all: one layer below the widening `ea0df4f` closed, in the same four verbs.
599///
600/// It now calls [`crate::gate::rule_hits`] — the gate's own firing decision — so the two cannot disagree
601/// again. That needs the function's TRANSITIVE reason classes and its ⟨0.20⟩ destination classes, the
602/// same two accumulators the gate reads, which is why they are parameters now rather than derivable.
603///
604/// **THE DIRECTION IS THE MIRROR ARGUMENT.** A filter can only ever SHRINK what a rule charges, so a
605/// filter-aware pass test can only ever find MORE holes — this cannot silently suppress a disclosure
606/// that used to appear. Pinned in both directions by
607/// `a_narrowed_rule_the_gate_tolerates_is_a_hole_and_the_one_it_fires_on_is_not`, in one run, because
608/// the fixture proving the fabrication is closed cannot show the reach closed with it. A WITHHELD filter
609/// — nothing to read — likewise counts as PASSING here: the gate did not clear the function, and an
610/// advisory note's fail-safe direction is to disclose.
611pub fn unverified_hole_rule<'a, S: AsRef<str>>(
612 name: &str,
613 effects: &[S],
614 reason_classes: Option<&BTreeSet<String>>,
615 net_classes: &[String],
616 rules: &'a [PolicyRule],
617) -> Option<&'a PolicyRule> {
618 if !effects.iter().any(|e| e.as_ref() == UNKNOWN) {
619 return None;
620 }
621 let effs: Vec<&str> = effects.iter().map(AsRef::as_ref).collect();
622 rules.iter().find(|r| {
623 // in the rule's scope (a scopeless rule governs the whole unit) …
624 if let Some(s) = &r.scope {
625 if !scope_matches(name, s) {
626 return false;
627 }
628 }
629 // … and PASSES it — the gate's own answer, narrowing filters and all. Empty `hits` IS passing.
630 crate::gate::rule_hits(r, &effs, reason_classes, net_classes).hits.is_empty()
631 })
632}
633
634/// Parse a CANDOR_POLICY file (SPEC §6.2). One rule per line; `#` comments and blanks ignored:
635///
636/// ```text
637/// deny Net Db domain # functions whose path contains segment "domain" must not perform Net or Db
638/// deny Exec # no function anywhere may perform Exec
639/// deny Unknown api # functions in "api" must be fully resolvable (forbid the unverifiable)
640/// pure parse # functions whose path contains segment "parse" must be effect-free
641/// allow Net in billing api.stripe.com
642/// forbid domain -> infra
643/// ```
644///
645/// In a `deny` rule, leading tokens that name a known effect (or `Unknown`) are forbidden; the FIRST
646/// non-effect token is the scope and ends the rule. A `deny` naming no known effect is dropped (it is
647/// NOT a `pure` rule). Malformed/unknown lines are ignored with a warning — never silently widened.
648/// The §6.2 token separator: ASCII whitespace ONLY (space/tab/CR/LF/VT/FF). `split_whitespace`/`trim`
649/// use Unicode `White_Space`, which would split a NBSP/ideographic space that Java drops — a gateless-
650/// green cross-engine divergence (adversarial DSL review). A non-ASCII space stays part of its token, so
651/// the rule is malformed and ignored, uniformly.
652fn is_ascii_ws(c: char) -> bool {
653 matches!(c, ' ' | '\t' | '\n' | '\x0b' | '\x0c' | '\r')
654}
655
656pub fn parse_policy(text: &str) -> ParsedPolicy {
657 parse_policy_impl(text, true, &std::collections::BTreeMap::new())
658}
659/// As [`parse_policy`] but with `.candor/config` `unknown-alias` definitions (⟨0.19⟩, SPEC §6.2): an
660/// `Unknown[<name>]` filter resolves a user-defined `<name>` to its reason classes. The gate + `parsepolicy`
661/// pass the discovered aliases (via [`parse_unknown_aliases`]); a config alias never changes what bare
662/// `deny E Unknown` means (always `Unknown[*]`), so the rule stays legible from the policy alone.
663pub fn parse_policy_with_aliases(text: &str, aliases: &std::collections::BTreeMap<String, std::collections::BTreeSet<ReasonClass>>) -> ParsedPolicy {
664 parse_policy_impl(text, true, aliases)
665}
666/// ⟨0.24⟩ [`parse_policy_with_aliases`] but SILENT — for a caller that must inspect
667/// [`ParsedPolicy::errors`] / [`ParsedPolicy::used_aliases`] BEFORE it parses for real (candor-scan
668/// refuses before touching the classifier's accumulators). Silent so the ordinary parse warnings are not
669/// printed twice on the same text.
670pub fn parse_policy_silent(
671 text: &str,
672 aliases: &std::collections::BTreeMap<String, BTreeSet<ReasonClass>>,
673) -> ParsedPolicy {
674 parse_policy_impl(text, false, aliases)
675}
676/// Same as [`parse_policy`] but SILENT about malformed rules — for a SECOND, advisory re-parse within the
677/// same run (candor-scan parses once for the gate check and again for the `unverified` disclosure), so the
678/// CI log doesn't print every "ignoring policy rule …" warning twice (#21). The first parse already warned.
679pub fn parse_policy_quiet(text: &str) -> ParsedPolicy {
680 parse_policy_impl(text, false, &std::collections::BTreeMap::new())
681}
682fn parse_policy_impl(text: &str, warn: bool, aliases: &std::collections::BTreeMap<String, std::collections::BTreeSet<ReasonClass>>) -> ParsedPolicy {
683 macro_rules! warn_ignore { ($($a:tt)*) => { if warn { eprintln!($($a)*); } } }
684 let mut out = ParsedPolicy::default();
685 // ⟨0.24⟩ Record a line the parser did not honour (SPEC §3.1 `195d45a`), on the ONE list `parsepolicy`
686 // reports and the gate routes filter for `fatal`. The stderr sentence and `message` are the SAME
687 // string by construction — a disclosure that can drift from the one beside it is how this family
688 // produced a FALSE disclosure once already (conformance PART 13b).
689 macro_rules! not_honoured {
690 ($fatal:expr, $kind:expr, $token:expr, $accepted:expr, $rule:expr, $msg:expr) => {{
691 let message: String = $msg;
692 out.errors.push(PolicyError {
693 kind: $kind,
694 token: ($token).to_string(),
695 accepted: ($accepted).iter().map(|s: &&str| s.to_string()).collect(),
696 rule: ($rule).to_string(),
697 message,
698 fatal: $fatal,
699 });
700 }};
701 }
702 // `str::lines()` splits on \n and \r\n but NOT bare \r — a classic-Mac file then collapses to ONE
703 // line, and since \r is also an in-line ASCII-ws token separator (is_ascii_ws), every rule after the
704 // first was glued into the first rule's tokens and dropped (sweep [16], a gateless-green divergence).
705 // Java's Files.readAllLines (the reference) breaks on bare \r too — normalize to match it. Allocation
706 // only when a bare \r is actually present (the overwhelmingly-common \n / \r\n files are untouched).
707 let normalized;
708 let text = if text.contains('\r') {
709 normalized = text.replace("\r\n", "\n").replace('\r', "\n");
710 normalized.as_str()
711 } else {
712 text
713 };
714 for raw_line in text.lines() {
715 let line = raw_line.split('#').next().unwrap_or("").trim_matches(is_ascii_ws);
716 if line.is_empty() {
717 continue;
718 }
719 let mut toks = line.split(is_ascii_ws).filter(|s| !s.is_empty());
720 match toks.next().unwrap_or("") {
721 "allow" => {
722 let effect = match toks.next().unwrap_or("") {
723 "Net" => "Net",
724 // `Llm` ⟨0.13⟩ rides the Net host literal (SPEC §1) — `allow Llm <host…>` restricts which
725 // MODEL hosts a scope may reach, matched by hostname like Net (its reached surface IS the
726 // Net host surface). Match candor-java's Policy.parsePolicy.
727 "Llm" => "Llm",
728 "Exec" => "Exec",
729 "Fs" => "Fs",
730 "Db" => "Db",
731 other => {
732 let msg = format!(
733 "unknown effect-name `{other}` in `allow` (accepted: Db, Exec, Fs, Llm, Net \
734 \u{2014} `allow` covers only the effects carrying a literal surface: Net/Llm \
735 hosts, Exec commands, Fs paths, Db tables): {line}"
736 );
737 warn_ignore!("candor: policy error — {msg}");
738 // ⟨0.24⟩ FATAL (SPEC §6.2 `1e1748a`). MEASURED four-way before this:
739 // `allow Nett host.example` -> exit 0 on rust, ts, java AND swift. The rule is
740 // DELETED and the certification silently vanishes, so the operator reads an
741 // armed allowlist that does not exist.
742 //
743 // The grammar defence that kept the token rule inside the bracket does NOT
744 // reach here: `allow`'s effect position is a fixed, closed set with **no scope
745 // reading available**, so an unrecognised token there is unambiguously a typo
746 // and there is no legitimate policy it could be. This document already calls a
747 // dropped rule "the limit case of silently rewritten into a different policy…
748 // a bigger rewrite than a narrowed filter" — and the bigger rewrite was
749 // warning-only while the smaller one was already exit 2.
750 not_honoured!(
751 true,
752 PolicyError::KIND_EFFECT_NAME,
753 other,
754 ["Db", "Exec", "Fs", "Llm", "Net"],
755 line,
756 msg
757 );
758 continue;
759 }
760 };
761 let mut rest: Vec<&str> = toks.collect();
762 let scope = if rest.first() == Some(&"in") {
763 let s = rest.get(1).map(|s| s.to_string());
764 rest.drain(..2.min(rest.len()));
765 s
766 } else {
767 None
768 };
769 let literals: BTreeSet<String> = rest.iter().map(|h| h.to_string()).collect();
770 if literals.is_empty() {
771 let msg = format!("`allow {effect}` names no values: {line}");
772 warn_ignore!("candor: ignoring policy rule ({msg})");
773 // `accepted` is EMPTY on purpose: the position takes an open-ended literal (a host, a
774 // path, a command, a table), so there is no token list to offer. A fact about the
775 // grammar, not a gap in the report.
776 not_honoured!(false, PolicyError::KIND_RULE_KIND, "", [], line, msg);
777 continue;
778 }
779 out.allow_rules.push(AllowRule { effect, scope, literals, raw: line.to_string() });
780 }
781 "deny" => {
782 let mut effects = BTreeSet::new();
783 let mut scope = None;
784 // Reason-class filter on `Unknown` (REASON-SCOPED-UNKNOWN-DESIGN.md): empty ⇒ `Unknown[*]`
785 // (any reason — the bare form); non-empty ⇒ only those classes. `*` = all.
786 let mut unknown_classes: BTreeSet<ReasonClass> = BTreeSet::new();
787 let mut unknown_star = false;
788 // Destination-class filter on `Net` (NET-DESTINATION-CLASS-DESIGN.md): empty ⇒ `Net[*]`
789 // (any destination — the bare form); non-empty ⇒ only those classes. `*` = all.
790 let mut net_classes: BTreeSet<String> = BTreeSet::new();
791 let mut net_star = false;
792 for t in toks {
793 // `Net[unknown-host]` / `Net[*]` / `Net[known-telemetry,unknown-host]`: the destination-scoped form.
794 if let Some(inner) = t.strip_prefix("Net[").and_then(|s| s.strip_suffix(']')) {
795 effects.insert("Net");
796 for cn in inner.split(',') {
797 let cn = cn.trim();
798 if cn.is_empty() {
799 continue;
800 }
801 if cn == "*" {
802 net_star = true;
803 } else if crate::NET_DEST_CLASSES.contains(&cn) {
804 net_classes.insert(cn.to_string());
805 } else {
806 // ⟨0.24⟩ A POLICY ERROR, not a warning — SPEC §6.2 `be0b9a9`. Byte-identical
807 // in shape to the reason-class arm below, and byte-identical in harm:
808 // MEASURED `deny Net[known-telemetry,unknown-hosst]` → exit 0 where the
809 // correctly-spelled rule exits 1. The typo is dropped, the filter NARROWS
810 // to `[known-telemetry]`, and the gate stops covering unidentifiable
811 // destinations while the operator reads a gate that looks armed.
812 not_honoured!(
813 true,
814 PolicyError::KIND_NET_CLASS,
815 cn,
816 ["known-telemetry", "known-partner", "unknown-host", "*"],
817 line,
818 format!(
819 "unrecognised Net destination-class `{cn}` in `{line}` — \
820 accepted: known-telemetry, known-partner, unknown-host, plus `*`"
821 )
822 );
823 }
824 }
825 continue;
826 }
827 // `Unknown[dispatch,reflect]` / `Unknown[*]` / `Unknown[dynamic]`: the reason-scoped form.
828 if let Some(inner) = t.strip_prefix("Unknown[").and_then(|s| s.strip_suffix(']')) {
829 effects.insert(UNKNOWN);
830 for cn in inner.split(',') {
831 let cn = cn.trim();
832 if cn.is_empty() {
833 continue;
834 }
835 if cn == "*" {
836 unknown_star = true;
837 } else if cn == "dynamic" {
838 unknown_classes.extend(ReasonClass::dynamic_set());
839 } else if let Some(rc) = ReasonClass::from_token(cn) {
840 unknown_classes.insert(rc);
841 } else if let Some(a) = aliases.get(cn) {
842 unknown_classes.extend(a.iter().copied()); // ⟨0.19⟩ config `unknown-alias`
843 // ⟨0.24⟩ → the verdict names it AND what it expanded to (SPEC §3.1
844 // `7f5b5ba`): the NAME alone cannot tell a reader which gate ran.
845 out.used_aliases
846 .insert(cn.to_string(), a.iter().map(|c| c.token().to_string()).collect());
847 } else {
848 // ⟨0.24⟩ A POLICY ERROR, not a warning — see `ParsedPolicy::errors`. The
849 // token is still dropped below so `rules` stays well-formed for the
850 // advisory readers (`unverified`, `parsepolicy`); the gate routes refuse
851 // on `errors` before any of it is used as a verdict.
852 not_honoured!(
853 true,
854 PolicyError::KIND_REASON_CLASS,
855 cn,
856 [
857 "reflect",
858 "dispatch",
859 "indirect",
860 "native",
861 "unresolved",
862 "setup",
863 "dynamic",
864 "*"
865 ],
866 line,
867 format!(
868 "unrecognised reason-class/alias `{cn}` in `{line}` — accepted: \
869 reflect, dispatch, indirect, native, unresolved, setup, plus the \
870 aliases `dynamic` and `*`, plus any `unknown-alias` defined in \
871 the `.candor/config` beside the policy. (⟨0.24⟩ an \
872 `unknown-alias` whose OWN definition names an unrecognised class \
873 is refused WHOLE, so a typo in the config surfaces as an \
874 undefined alias here — check the `unknown-alias` lines too, and \
875 the line above this one.)"
876 )
877 );
878 }
879 }
880 continue;
881 }
882 let e = if t == UNKNOWN { Some(UNKNOWN) } else { cap_from_name(t) };
883 match e {
884 Some(e) => {
885 effects.insert(e);
886 if e == UNKNOWN {
887 unknown_star = true; // bare Unknown ⇒ all classes
888 }
889 if e == "Net" {
890 net_star = true; // bare Net ⇒ all destinations
891 }
892 }
893 None => {
894 scope = Some(t.to_string());
895 break;
896 }
897 }
898 }
899 if effects.is_empty() {
900 // The accepted set is the §1 effect vocabulary plus `Unknown` — SORTED, so the
901 // document is deterministic and diffable across engines.
902 let mut acc: Vec<&str> = candor_report::EFFECTS.to_vec();
903 acc.push(UNKNOWN);
904 acc.sort_unstable();
905 let msg = format!(
906 "`deny` names no known effect (accepted: {}): {line}",
907 acc.join(", ")
908 );
909 warn_ignore!("candor: policy error — {msg}");
910 // ⟨0.24⟩ FATAL (SPEC §6.2 `1e1748a`). MEASURED four-way: `deny Nett app` -> exit 0
911 // on all four; the rule is DELETED and the gate is green. `Nett` is read as the
912 // SCOPE (the first unrecognised token ends the effect list), so the line parses to
913 // a deny of NOTHING.
914 //
915 // **A `deny` whose effect list ends up EMPTY is malformed under EITHER reading** —
916 // typo-in-the-effect or scope-with-no-effect are both nonsense — so there is no
917 // legitimate policy it could be and refusing it loses nothing. What stays open is
918 // only the genuinely ambiguous middle (`deny Net Exex app`: at least one valid
919 // effect plus an unrecognised trailing token that MIGHT be a scope), which the
920 // parser cannot tell from a legitimate scope and which `parsepolicy` shows either
921 // way by dumping the `scope` it read.
922 not_honoured!(
923 true,
924 PolicyError::KIND_EFFECT_NAME,
925 scope.as_deref().unwrap_or(""),
926 acc,
927 line,
928 msg
929 );
930 continue;
931 }
932 // `*` (or bare `Unknown`) means all classes ⇒ empty filter (matches any Unknown).
933 if unknown_star {
934 unknown_classes.clear();
935 } else if !unknown_classes.is_empty() && !unknown_classes.contains(&ReasonClass::Unresolved) {
936 // A2 under-gating lint: a narrowed scope that omits `unresolved` (the catch-all for holes
937 // an engine couldn't classify) may silently tolerate exactly those — flag it (advisory).
938 warn_ignore!("candor: policy rule narrows `Unknown[…]` but omits `unresolved` — may UNDER-gate on holes the engine couldn't classify; add `unresolved` (or use `dynamic`) to stay conservative: {line}");
939 }
940 // `*` (or bare `Net`) means all destinations ⇒ empty filter (matches any Net).
941 if net_star {
942 net_classes.clear();
943 }
944 out.rules.push(PolicyRule { effects, scope, unknown_classes, net_classes, raw: line.to_string() });
945 }
946 "pure" => out.rules.push(PolicyRule {
947 effects: BTreeSet::new(),
948 scope: toks.next().map(str::to_string),
949 unknown_classes: BTreeSet::new(),
950 net_classes: BTreeSet::new(),
951 raw: line.to_string(),
952 }),
953 "forbid" => {
954 let a = toks.next().unwrap_or("");
955 let arrow = toks.next().unwrap_or("");
956 let b = toks.next().unwrap_or("");
957 if a.is_empty() || arrow != "->" || b.is_empty() {
958 let msg = format!("`forbid` is malformed (want `forbid <scope> -> <scope>`): {line}");
959 warn_ignore!("candor: ignoring layering rule ({msg})");
960 // The token reported is whatever sat in the ARROW position — `->` must be its own
961 // token, so `forbid glued->arrow` finds nothing there and that absence is the finding.
962 not_honoured!(false, PolicyError::KIND_RULE_KIND, arrow, ["->"], line, msg);
963 continue;
964 }
965 out.layer_rules.push(LayerRule {
966 from: a.to_string(),
967 to: b.to_string(),
968 raw: line.to_string(),
969 });
970 }
971 other => {
972 let msg = format!(
973 "unknown rule kind `{other}` (accepted: allow, deny, forbid, pure): {line}"
974 );
975 warn_ignore!("candor: ignoring policy rule ({msg})");
976 not_honoured!(
977 false,
978 PolicyError::KIND_RULE_KIND,
979 other,
980 ["allow", "deny", "forbid", "pure"],
981 line,
982 msg
983 );
984 }
985 }
986 }
987 out
988}
989
990#[cfg(test)]
991mod tests {
992 #[test]
993 fn db_table_covering_is_strict() {
994 use super::db_table_covered as c;
995 assert!(c("ledger.entries", "Ledger.Entries")); // case-insensitive exact
996 assert!(c("ledger.*", "ledger.entries")); // schema wildcard
997 assert!(!c("ledger.*", "ledgerx.entries")); // boundary-respecting
998 assert!(!c("entries", "ledger.entries")); // no silent qualification widening
999 assert!(c("entries", "entries"));
1000 }
1001
1002 #[test]
1003 fn allow_db_parses_and_gates() {
1004 let p = super::parse_policy("allow Db in billing ledger.* customers\n");
1005 assert_eq!(p.allow_rules.len(), 1);
1006 assert_eq!(p.allow_rules[0].effect, "Db");
1007 assert!(super::literal_allowed("Db", "ledger.entries", &p.allow_rules[0].literals));
1008 assert!(super::literal_allowed("Db", "customers", &p.allow_rules[0].literals));
1009 assert!(!super::literal_allowed("Db", "audit.log", &p.allow_rules[0].literals));
1010 }
1011
1012 use super::*;
1013
1014 #[test]
1015 fn policy_parses() {
1016 let p = parse_policy(
1017 "# the domain layer must stay pure of I/O\n\
1018 deny Net Db domain\n\
1019 deny Exec\n\
1020 pure parse\n\
1021 nonsense line\n\
1022 deny notaneffect\n",
1023 );
1024 let rules = &p.rules;
1025 assert_eq!(rules.len(), 3);
1026 assert_eq!(rules[0].effects, ["Db", "Net"].into_iter().collect::<BTreeSet<_>>());
1027 assert_eq!(rules[0].scope.as_deref(), Some("domain"));
1028 assert!(rules[1].effects.contains("Exec") && rules[1].scope.is_none());
1029 assert!(rules[2].effects.is_empty() && rules[2].scope.as_deref() == Some("parse"));
1030 // sweep [16]: a classic-Mac (bare \r) multi-rule policy must NOT collapse to the first rule.
1031 let cr = parse_policy("deny Net a\rdeny Exec b\rdeny Db c\r");
1032 assert_eq!(cr.rules.len(), 3, "bare-CR lines must each parse");
1033 assert!(cr.rules.iter().any(|r| r.effects.contains("Exec") && r.scope.as_deref() == Some("b")));
1034 // mixed \r\n and bare \r normalize identically.
1035 assert_eq!(parse_policy("deny Net a\r\ndeny Exec b\r").rules.len(), 2);
1036 // `Unknown` is a denyable token; a bare `deny` with no effect is ignored.
1037 assert_eq!(parse_policy("deny Unknown core").rules[0].effects, ["Unknown"].into_iter().collect());
1038 assert!(parse_policy("deny\ndeny \n").rules.is_empty());
1039 // a `deny` whose first token is a non-effect names no effect -> dropped, NOT a pure rule.
1040 assert!(parse_policy("deny notaneffect scope").rules.is_empty());
1041 // the first non-effect token ENDS the rule: a later effect token is not collected.
1042 let p2 = parse_policy("deny Net foo Db");
1043 assert_eq!(p2.rules[0].effects, ["Net"].into_iter().collect::<BTreeSet<_>>());
1044 assert_eq!(p2.rules[0].scope.as_deref(), Some("foo"));
1045 // NBSP is NOT a token separator (only ASCII White_Space is) — pinned to MATCH Java, which
1046 // drops it: a `deny\u{a0}Net` is one token `deny\u{a0}Net`, NOT `deny` + `Net`, so it names no
1047 // known effect and is dropped. Splitting on Unicode whitespace here would let candor see a deny
1048 // the JVM engine doesn't — a gateless-divergence between impls. (See is_ascii_ws.)
1049 assert!(parse_policy("deny\u{a0}Net core").rules.is_empty(),
1050 "an NBSP between deny and the effect must NOT split into separate tokens");
1051 // The NBSP rides INTO the scope token rather than separating it: `deny Net\u{a0}domain` is
1052 // `deny` + `Net` + `\u{a0}domain` — Net is the effect, the scope keeps the NBSP verbatim.
1053 let nb = parse_policy("deny Net \u{a0}domain");
1054 assert_eq!(nb.rules.len(), 1);
1055 assert_eq!(nb.rules[0].effects, ["Net"].into_iter().collect::<BTreeSet<_>>());
1056 assert_eq!(nb.rules[0].scope.as_deref(), Some("\u{a0}domain"));
1057 }
1058
1059 #[test]
1060 fn reason_scoped_unknown_parses() {
1061 use super::ReasonClass::*;
1062 // `Unknown[dispatch,indirect]` narrows the Unknown membership to those classes.
1063 let p = parse_policy("deny Net Unknown[dispatch,indirect] dom\n");
1064 let r = &p.rules[0];
1065 assert!(r.effects.contains("Unknown") && r.effects.contains("Net"));
1066 assert_eq!(r.scope.as_deref(), Some("dom"));
1067 assert_eq!(r.unknown_classes, [Dispatch, Indirect].into_iter().collect());
1068 // bare `Unknown` and `Unknown[*]` ⇒ empty filter (all classes).
1069 assert!(parse_policy("deny Net Unknown dom\n").rules[0].unknown_classes.is_empty(), "bare Unknown ⇒ all");
1070 assert!(parse_policy("deny Net Unknown[*] dom\n").rules[0].unknown_classes.is_empty(), "Unknown[*] ⇒ all");
1071 // `dynamic` alias = every genuine class incl. unresolved, excl. setup.
1072 assert_eq!(
1073 parse_policy("deny Net Unknown[dynamic] dom\n").rules[0].unknown_classes,
1074 [Reflect, Dispatch, Indirect, Native, Unresolved].into_iter().collect()
1075 );
1076 // config `unknown-alias` (⟨0.19⟩): a user-defined name resolves; a reserved name is rejected.
1077 let aliases = super::parse_unknown_aliases(
1078 "unknown-alias risky = reflect,native\nunknown-alias telemetry = indirect\nunknown-alias reflect = native\n");
1079 assert_eq!(aliases.get("risky"), Some(&[Reflect, Native].into_iter().collect()));
1080 assert_eq!(aliases.get("telemetry"), Some(&[Indirect].into_iter().collect()));
1081 assert!(!aliases.contains_key("reflect"), "a config alias may not shadow a class token");
1082 // the `unknown-alias` KEY matches case-insensitively (parity with java/ts/swift, which lowercase it)
1083 assert_eq!(super::parse_unknown_aliases("Unknown-Alias hot = native\n").get("hot"),
1084 Some(&[Native].into_iter().collect()), "the unknown-alias key must match case-insensitively");
1085 let pr = super::parse_policy_with_aliases("deny Net Unknown[risky] api\n", &aliases);
1086 assert_eq!(pr.rules[0].unknown_classes, [Reflect, Native].into_iter().collect());
1087 // an UNDEFINED alias name is dropped-with-warning → empty filter (behaves like bare Unknown[*])
1088 assert!(super::parse_policy_with_aliases("deny Net Unknown[nope] api\n", &aliases).rules[0].unknown_classes.is_empty());
1089 // classify: raw reason tokens → normative classes (mirrors java ReasonClass.classify).
1090 assert_eq!(ReasonClass::classify("reflect:Class.forName"), Reflect);
1091 assert_eq!(ReasonClass::classify("native:extern fn"), Native);
1092 assert_eq!(ReasonClass::classify("callback:unresolved call"), Indirect);
1093 assert_eq!(ReasonClass::classify("ambiguous:same-name local defs"), Dispatch);
1094 assert_eq!(ReasonClass::classify("unresolved"), Unresolved);
1095 assert_eq!(ReasonClass::classify("whatever-new"), Unresolved); // conservative catch-all
1096 }
1097
1098 /// THE CONTROL SPEC §4 ⟨0.24⟩ MAKES A SHOULD: a FABRICATED, off-vocabulary kind must still behave as
1099 /// §2 forward-compatibility requires. Without it, "added a fifth kind" and "stopped checking the kind
1100 /// set" are the same diff — the classifier is one `_ =>` arm away from either.
1101 ///
1102 /// This engine holds the §4 vocabulary ONCE (the raw `kind:detail` string, read back only through
1103 /// `classify`), so there is no typed half here to drift from it. That is why the JVM engine's failure
1104 /// — a string classifier correct on `ambiguous` since July while its typed `Kind` enum lacked the kind
1105 /// entirely, one token classified two ways inside one engine — is not reproducible here. If a typed
1106 /// kind representation is ever added, this test is where its half gets its control.
1107 #[test]
1108 fn off_vocabulary_kinds_round_trip_and_classify_through_the_catch_all() {
1109 use ReasonClass::*;
1110 // A kind no engine emits and no section names. §2: tolerated, and classified CONSERVATIVELY —
1111 // `unresolved`, the catch-all, so a narrowed `Unknown[unresolved]`/`[dynamic]`/`[*]` still bites it.
1112 assert_eq!(ReasonClass::classify("banana:whatever"), Unresolved);
1113 assert_eq!(ReasonClass::classify("banana:dispatch of a banana"), Unresolved,
1114 "a canonical kind appearing in the DETAIL must not leak into the classification");
1115 // …and it must not be swallowed into a narrower class. These are the four wrong answers.
1116 for wrong in [Reflect, Dispatch, Indirect, Native] {
1117 assert_ne!(ReasonClass::classify("banana:whatever"), wrong);
1118 }
1119 // The five §4 ⟨0.24⟩ kinds all classify, and `ambiguous` is the fifth — pinned beside the
1120 // fabricated one deliberately: one arm chain answers both, so a change that stops distinguishing
1121 // them fails here rather than in the field.
1122 assert_eq!(ReasonClass::classify("reflect:x"), Reflect);
1123 assert_eq!(ReasonClass::classify("native:x"), Native);
1124 assert_eq!(ReasonClass::classify("dispatch:Owner.member"), Dispatch);
1125 assert_eq!(ReasonClass::classify("callback:x"), Indirect);
1126 assert_eq!(ReasonClass::classify("ambiguous:x"), Dispatch);
1127 // ⟨0.24⟩ `dep:<hash>` / `dep-stale:<pkg>` are REGISTERED §4 kinds, not migration ones — swift
1128 // emits them per dependency ENTRY, and this engine CONSUMES swift/ts reports through
1129 // candor-query. §6.2 pins their class as `unresolved`, which is where the catch-all lands them;
1130 // pinned so a future prefix arm cannot move them without saying so.
1131 assert_eq!(ReasonClass::classify("dep:9f2c1a"), Unresolved);
1132 assert_eq!(ReasonClass::classify("dep-stale:somepkg"), Unresolved);
1133 }
1134
1135 #[test]
1136 fn net_destination_class_parses_and_classifies() {
1137 // `Net[unknown-host,known-telemetry]` narrows the Net membership to those destination classes.
1138 let p = parse_policy("deny Net[unknown-host,known-telemetry] dom\n");
1139 let r = &p.rules[0];
1140 assert!(r.effects.contains("Net"));
1141 assert_eq!(r.scope.as_deref(), Some("dom"));
1142 assert_eq!(
1143 r.net_classes,
1144 ["unknown-host", "known-telemetry"].iter().map(|s| s.to_string()).collect()
1145 );
1146 // bare `Net` and `Net[*]` ⇒ empty filter (all destinations).
1147 assert!(parse_policy("deny Net dom\n").rules[0].net_classes.is_empty(), "bare Net ⇒ all");
1148 assert!(parse_policy("deny Net[*] dom\n").rules[0].net_classes.is_empty(), "Net[*] ⇒ all");
1149 // an unknown destination-class is dropped-with-warning → empty filter (behaves like bare Net[*]).
1150 assert!(parse_policy("deny Net[nope] dom\n").rules[0].net_classes.is_empty());
1151 // the classifier: telemetry (subdomain-aware), model host, unresolved, and the config-partner path.
1152 let no_partners = BTreeSet::new();
1153 assert_eq!(crate::net_dest_class("sentry.io", &no_partners), "known-telemetry");
1154 assert_eq!(crate::net_dest_class("us.i.posthog.com", &no_partners), "known-telemetry"); // 0.20.1 corpus-grown
1155 assert_eq!(crate::net_dest_class("o1.ingest.sentry.io", &no_partners), "known-telemetry");
1156 assert_eq!(crate::net_dest_class("api.openai.com", &no_partners), "known-partner", "a model host is known-partner");
1157 assert_eq!(crate::net_dest_class("evil.example.com", &no_partners), "unknown-host");
1158 let partners: BTreeSet<String> = ["api.stripe.com".to_string()].into_iter().collect();
1159 assert_eq!(crate::net_dest_class("api.stripe.com", &partners), "known-partner", "config-declared partner");
1160 assert_eq!(crate::net_dest_class("api.stripe.com", &no_partners), "unknown-host", "partner is config-only");
1161 // `net-partner` config parsing: host-normalized, case-insensitive key, multi-value.
1162 let pset = super::parse_net_partners("net-partner Api.Stripe.com:443\nNET-PARTNER hooks.stripe.com\n");
1163 assert!(pset.contains("api.stripe.com") && pset.contains("hooks.stripe.com"));
1164 }
1165
1166 #[test]
1167 fn allowlist_parses() {
1168 let p = parse_policy(
1169 "allow Net in billing api.stripe.com hooks.stripe.com\n\
1170 allow Exec in ci git\n\
1171 allow Fs in config /etc/app\n\
1172 allow Net github.com\n\
1173 allow Clock whatever\n\
1174 allow Net in nohosts\n\
1175 allow\n",
1176 );
1177 assert_eq!(p.allow_rules.len(), 4); // Clock carries no literal surface — rejected; Db now does
1178 assert_eq!((p.allow_rules[0].effect, p.allow_rules[0].scope.as_deref()), ("Net", Some("billing")));
1179 assert_eq!(
1180 p.allow_rules[0].literals,
1181 ["api.stripe.com", "hooks.stripe.com"].iter().map(|s| s.to_string()).collect()
1182 );
1183 assert_eq!((p.allow_rules[1].effect, p.allow_rules[1].scope.as_deref()), ("Exec", Some("ci")));
1184 assert!(p.allow_rules[1].literals.contains("git"));
1185 assert_eq!((p.allow_rules[2].effect, p.allow_rules[2].scope.as_deref()), ("Fs", Some("config")));
1186 assert_eq!((p.allow_rules[3].effect, p.allow_rules[3].scope.is_none()), ("Net", true));
1187
1188 let set = |xs: &[&str]| xs.iter().map(|s| s.to_string()).collect::<BTreeSet<_>>();
1189 assert!(literal_allowed("Net", "api.stripe.com:443", &set(&["api.stripe.com"])));
1190 // IPv6: a bare literal is matched WHOLE (no first-colon collapse), so a different address in the
1191 // same block is NOT accepted; a bracketed `[host]:port` matches the bare host. (/code-review.)
1192 assert!(literal_allowed("Net", "2001:db8::aa", &set(&["2001:db8::aa"])));
1193 assert!(!literal_allowed("Net", "2001:db8::ff", &set(&["2001:db8::aa"])));
1194 assert!(!literal_allowed("Net", "2001:dead::1", &set(&["2001:db8::aa"])));
1195 assert!(literal_allowed("Net", "[2001:db8::aa]:443", &set(&["2001:db8::aa"])));
1196 assert_eq!(host_part("2001:db8::aa"), "2001:db8::aa");
1197 assert_eq!(host_part("[2001:db8::aa]:443"), "2001:db8::aa");
1198 assert_eq!(host_part("api.stripe.com:443"), "api.stripe.com");
1199 assert!(literal_allowed("Exec", "/usr/bin/git", &set(&["git"])));
1200 assert!(!literal_allowed("Exec", "/usr/bin/curl", &set(&["git"])));
1201 assert!(literal_allowed("Fs", "/etc/app/conf.toml", &set(&["/etc/app"])));
1202 assert!(!literal_allowed("Fs", "/etc/shadow", &set(&["/etc/app"])));
1203 assert_eq!(cmd_base("/usr/bin/git"), "git");
1204 }
1205
1206 #[test]
1207 fn layering_rule_parses() {
1208 let p = parse_policy(
1209 "forbid domain -> infra\n\
1210 forbid app::web -> app::db \n\
1211 forbid domain infra\n\
1212 forbid domain ->\n\
1213 forbid\n",
1214 );
1215 assert_eq!(p.layer_rules.len(), 2);
1216 assert_eq!((p.layer_rules[0].from.as_str(), p.layer_rules[0].to.as_str()), ("domain", "infra"));
1217 assert_eq!((p.layer_rules[1].from.as_str(), p.layer_rules[1].to.as_str()), ("app::web", "app::db"));
1218 }
1219
1220 #[test]
1221 fn scope_matches_by_segment_not_substring() {
1222 assert!(scope_matches("app::domain::handle", "domain"));
1223 assert!(scope_matches("domain::handle", "domain"));
1224 assert!(scope_matches("app::domain", "domain"));
1225 assert!(scope_matches("crate::domain_logic", "domain"));
1226 assert!(!scope_matches("app::subdomain::handle", "domain"));
1227 assert!(!scope_matches("app::not_my_domain::f", "domain"));
1228 // multi-segment: intermediates exact, last is a prefix, contiguous.
1229 assert!(scope_matches("crate::net::client::send", "net::client"));
1230 assert!(scope_matches("crate::net::client_pool::get", "net::client"));
1231 assert!(!scope_matches("crate::net::server::send", "net::client"));
1232 assert!(!scope_matches("crate::network::client::send", "net::client"));
1233 assert!(!scope_matches("crate::net::x::client", "net::client"));
1234 assert!(!scope_matches("net", "net::client"));
1235 // DOTTED names (JVM/Swift/TS reports `candor-query` consumes): a scope must match across `.` too,
1236 // else a scoped deny/pure rule is silently inert → whatif false-green (gate-evasion). Both a
1237 // `.`-written and a `::`-written scope must match a dotted name.
1238 assert!(scope_matches("com.acme.domain.Pricing.quote", "domain"));
1239 assert!(scope_matches("com.acme.domain.Pricing.quote", "acme.domain"));
1240 assert!(scope_matches("com.acme.domain.Pricing.quote", "acme::domain"));
1241 assert!(scope_matches("com.acme.infra.Net.fetch", "infra.Net"));
1242 assert!(!scope_matches("com.acme.subdomain.h", "domain"));
1243 assert!(!scope_matches("com.acme.domain.h", "infra"));
1244 }
1245
1246 #[test]
1247 fn fs_path_covered_respects_boundaries() {
1248 assert!(fs_path_covered("/etc/app", "/etc/app"));
1249 assert!(fs_path_covered("/etc/app", "/etc/app/cfg.toml"));
1250 assert!(fs_path_covered("/etc/app/", "/etc/app/cfg"));
1251 assert!(!fs_path_covered("/etc/app", "/etc/apppwned"));
1252 assert!(!fs_path_covered("/etc/app", "/etc/application/x"));
1253 assert!(!fs_path_covered("/etc/app/cfg", "/etc/app"));
1254 assert!(!fs_path_covered("/etc/app", "/etc/app/../passwd"));
1255 assert!(fs_path_covered("/", "/etc/app/x"));
1256 assert!(!fs_path_covered("etc/app", "/etc/app/cfg"));
1257 assert!(!fs_path_covered("/etc/app", "etc/app/cfg"));
1258 assert!(fs_path_covered("etc/app", "etc/app/cfg"));
1259 }
1260
1261 /// ⟨0.24⟩ THE PROVABLE-PURITY DISCLOSURE MUST ASK THE GATE WHAT "PASSES" MEANS — SPEC §6.2, and both
1262 /// directions in ONE test, because killing an over-charge is exactly where a silent under-report gets
1263 /// introduced and the fixture proving the fabrication is closed cannot show the reach closed with it.
1264 ///
1265 /// A hole is a function that PASSES its rule while `Unknown`. `unverified_hole_rule` used to compute
1266 /// PASSES from `r.effects` alone — the pre-⟨0.19⟩ question, asked after two rungs gave rules a
1267 /// NARROWING FILTER — so a rule the gate TOLERATES was read here as violated and the hole was DELETED
1268 /// from the disclosure. MEASURED 2026-07-28: `deny Unknown[reflect]` over an `indirect` hole → gate
1269 /// exit 0, `unverified` "every function in a pure/deny layer is PROVABLY clean ✓".
1270 ///
1271 /// ROW 1 (the fix) — the filter does NOT match, so the gate tolerates and this IS a hole.
1272 /// ROW 2 (the mirror) — the SAME rule, SAME function, filter spelled to MATCH: the gate fires, so it
1273 /// is a violation and NOT a hole. Without row 2 the fix is satisfied by a predicate that calls
1274 /// everything a hole, which is the mirror over-report of the thing being fixed.
1275 /// ROW 3 — the ⟨0.20⟩ `Net[dest…]` filter, the same shape on the other narrowing axis.
1276 /// ROW 4 — no filter at all: byte-identical to pre-⟨0.24⟩, which is what keeps conformance PARTs
1277 /// 12c/12d (four-way) from moving.
1278 #[test]
1279 fn a_narrowed_rule_the_gate_tolerates_is_a_hole_and_the_one_it_fires_on_is_not() {
1280 let effs = ["Unknown"];
1281 let indirect: BTreeSet<String> = ["indirect".to_string()].into_iter().collect();
1282 let hole = |src: &str, classes: Option<&BTreeSet<String>>, nets: &[String]| -> Option<String> {
1283 let p = parse_policy(src);
1284 unverified_hole_rule("app::go", &effs, classes, nets, &p.rules).map(|r| rule_and_upgrade(r).1)
1285 };
1286
1287 // ROW 1 — tolerated by the gate (indirect ∉ {reflect}) ⇒ a hole, and the upgrade WIDENS the
1288 // filter rather than appending a second `Unknown`.
1289 assert_eq!(
1290 hole("deny Unknown[reflect]\n", Some(&indirect), &[]).as_deref(),
1291 Some("deny Unknown"),
1292 "a rule the gate TOLERATES leaves the function unproven — that is the disclosure's whole subject"
1293 );
1294
1295 // ROW 2 — THE MIRROR. Same rule, same signature, filter spelled to match: the gate FIRES, so this
1296 // is a violation and the disclosure must stay silent about it.
1297 assert_eq!(
1298 hole("deny Unknown[indirect]\n", Some(&indirect), &[]),
1299 None,
1300 "a rule the gate FIRES on is a violation, not a hole — filter-awareness must not start \
1301 disclosing the gate's own findings back as unproven passes"
1302 );
1303 assert_eq!(hole("deny Unknown[dynamic]\n", Some(&indirect), &[]), None, "`dynamic` covers indirect");
1304
1305 // ROW 3 — the ⟨0.20⟩ destination filter, both ways, on a fn carrying Net BESIDE its Unknown.
1306 let netfn = ["Net".to_string(), "Unknown".to_string()];
1307 let telemetry = vec!["known-telemetry".to_string()];
1308 let p = parse_policy("deny Net[unknown-host]\n");
1309 assert_eq!(
1310 unverified_hole_rule("app::go", &netfn, Some(&indirect), &telemetry, &p.rules)
1311 .map(|r| rule_and_upgrade(r).1)
1312 .as_deref(),
1313 Some("deny Net[unknown-host] Unknown"),
1314 "a `Net[dest…]` the fn's destinations do not match is tolerated, so the Unknown beside it is a hole"
1315 );
1316 let p = parse_policy("deny Net[known-telemetry]\n");
1317 assert_eq!(
1318 unverified_hole_rule("app::go", &netfn, Some(&indirect), &telemetry, &p.rules).map(|r| r.raw.clone()),
1319 None,
1320 "MIRROR: the matching destination filter FIRES, so it is a violation and not a hole"
1321 );
1322
1323 // ROW 4 — UNFILTERED, unchanged. `deny Unknown` fires on every Unknown (never a hole); `pure` and
1324 // `deny Fs` pass a fn with no real effect (always a hole) — the forms PARTs 12c/12d pin four-way.
1325 assert_eq!(hole("deny Unknown\n", Some(&indirect), &[]), None);
1326 assert_eq!(hole("pure\n", Some(&indirect), &[]).as_deref(), Some("deny Unknown"));
1327 assert_eq!(hole("deny Net Db domain\ndeny Fs\n", Some(&indirect), &[]).as_deref(), Some("deny Fs Unknown"));
1328 assert_eq!(
1329 hole("deny Net Db go\n", Some(&indirect), &[]).as_deref(),
1330 Some("deny Db Net Unknown go"),
1331 "the sorted multi-effect upgrade PART 12c-deny pins in all four engines"
1332 );
1333
1334 // A WITHHELD filter — no class set to read — counts as PASSING, so the hole is disclosed rather
1335 // than dropped. The gate withholds there too; between an advisory note that speaks and one that
1336 // goes quiet over a rule that never ran, only the first stays true.
1337 assert_eq!(hole("deny Unknown[reflect]\n", None, &[]).as_deref(), Some("deny Unknown"));
1338 }
1339}