candor_classify/gate.rs
1//! ⟨0.24⟩ THE GATE — SPEC §6.2 matching over an ALREADY-ACCUMULATED signature, and the ONLY copy of
2//! that matching in the stable toolchain.
3//!
4//! **THE SEAM.** [`GateInput`] is the boundary between *what produced the signature* and *what §6.2
5//! does with it*. Every field is already accumulated: this module runs no fixpoint, opens no file and
6//! consults no scan state, so the same matching code serves both routes in —
7//!
8//! - `candor-scan … --policy <f>` builds a `GateInput` from the classifier's transitive accumulators
9//! (`crate::gate::policy_violations`, which is now a thin wrapper);
10//! - `candor-query gate --report <loc> --policy <f>` (SPEC §3.1 ⟨0.24⟩) builds one from a WRITTEN
11//! report and nothing else.
12//!
13//! That split is the whole point of §3.1 ⟨0.24⟩: until it existed the gate was reachable only THROUGH
14//! the classifier, so a defect in the gate and a defect in the classifier were indistinguishable from
15//! any test that could be written. Do NOT re-implement the matching on the report side — the §6.2
16//! clause that mandates the verb was written about exactly that mistake.
17
18use crate::policy::{literal_allowed, reason_class_matches, scope_matches, scope_matches_permitted,
19 ParsedPolicy, PolicyRule};
20use candor_report::GateViolation;
21use std::collections::{BTreeSet, HashMap};
22
23/// ⟨0.20⟩ The `Net` destination classes an fn reaches (transitive) — the SINGLE derivation shared by the
24/// report's `netClass` field (candor-scan's writer) and the gate: an exact host-literal match
25/// ([`crate::net_dest_class`]) for the visible hosts, plus the fail-closed `unknown-host` when the Net
26/// surface is masked (`incomplete` has Net) OR carries no visible host (a runtime endpoint). Call only
27/// for an fn known to have Net; returns sorted.
28///
29/// It lives beside the gate rather than in the scanner because the report field and the gate filter MUST
30/// be the same set: `gate --report` reads `netClass` off the wire and the scan derives it here, and the
31/// §3.1 ⟨0.24⟩ byte-equivalence obligation is exactly the claim that those two agree.
32pub fn net_classes_of<E: AsRef<str> + Ord>(
33 q: &str,
34 hostsacc: &HashMap<String, BTreeSet<String>>,
35 incompleteacc: &HashMap<String, BTreeSet<E>>,
36 partners: &BTreeSet<String>,
37) -> Vec<String> {
38 let mut classes: BTreeSet<String> = hostsacc
39 .get(q)
40 .into_iter()
41 .flatten()
42 .map(|h| crate::net_dest_class(h, partners).to_string())
43 .collect();
44 let masked = incompleteacc.get(q).is_some_and(|s| s.iter().any(|e| e.as_ref() == "Net"));
45 let no_hosts = hostsacc.get(q).map(|s| s.is_empty()).unwrap_or(true);
46 if masked || no_hosts {
47 classes.insert("unknown-host".to_string());
48 }
49 classes.into_iter().collect()
50}
51
52/// ⟨0.24⟩ THE GATE'S INPUT — one signature per function, every field already TRANSITIVE.
53///
54/// `E` is the effect-name representation: `&'static str` on the scan route (the classifier's interned
55/// vocabulary) and `String` on the report route (the wire's names, taken VERBATIM — a report naming an
56/// effect this build's vocabulary does not list must still trip a `pure` rule, so the names are never
57/// filtered through a known-effect allowlist on the way in).
58impl<'a, E: AsRef<str> + Ord> GateInput<'a, E> {
59 /// The name to match a policy scope against, and to print. Identity when no map was supplied.
60 pub fn disp<'x>(&'x self, k: &'x str) -> &'x str {
61 self.display.get(k).map(|v| v.as_str()).unwrap_or(k)
62 }
63
64 /// ⟨0.32⟩ The §2.2 UNIT IDENTITY this key stands for — what a verdict row carries so a consumer can
65 /// tell two units apart (SPEC §2). EMPTY when the caller has none to give, and empty is then omitted
66 /// from the wire: *this producer cannot answer* beats a fabricated id.
67 ///
68 /// A separate map from [`Self::display`] rather than a second use of the key, because the two routes
69 /// disagree about what the key IS: the report route keys by `hash` already, the scan route keys by
70 /// the qualified NAME and must qualify it with the crate. Reading identity off the key would have
71 /// made the two routes emit different `hash` values for one unit, which §3.1 byte-equality forbids
72 /// and which no single-route test could see.
73 pub fn unit<'x>(&'x self, k: &'x str) -> &'x str {
74 self.hash.get(k).map(|v| v.as_str()).unwrap_or("")
75 }
76}
77
78pub struct GateInput<'a, E: AsRef<str> + Ord> {
79 /// Every UNIT the gate ranges over, in the caller's order — an opaque KEY, not necessarily a name.
80 pub all: &'a [String],
81 /// ⟨0.32⟩ key -> the name to MATCH and DISPLAY. Empty means the keys are already names.
82 ///
83 /// A multi-report gate must join by `hash` and never by bare `fn` (SPEC §2.2), because two members
84 /// of a workspace legitimately share a name — and merging them was measured turning a refusal into
85 /// `policy ✓` by letting one borrow the other's Unknown reason class. But a hash is `package#fn`,
86 /// and a POLICY SCOPE is written against the name (`deny Exec app::`), so keying by hash without
87 /// this map silently stops scopes matching: a false green introduced by fixing a false green.
88 /// Identity is the default, so the single-report callers are unaffected.
89 pub display: &'a std::collections::HashMap<String, String>,
90 /// ⟨0.32⟩ key -> the §2.2 UNIT IDENTITY (`package#fn`) a verdict row carries. Missing means this
91 /// caller has none, and the row then omits the field — see [`Self::unit`].
92 pub hash: &'a std::collections::HashMap<String, String>,
93 /// Per fn, the TRANSITIVE effect set — the model's `S`, with candor's `Unknown` marker carried as a
94 /// member (this engine's encoding of `D ≠ ∅`).
95 pub inferred: &'a HashMap<String, BTreeSet<E>>,
96 /// The call graph AS-EFF-009 walks.
97 pub calls: &'a HashMap<String, BTreeSet<String>>,
98 /// Per fn, the TRANSITIVE literal surface AS-EFF-008 certifies against.
99 pub hosts: &'a HashMap<String, BTreeSet<String>>,
100 pub cmds: &'a HashMap<String, BTreeSet<String>>,
101 pub paths: &'a HashMap<String, BTreeSet<String>>,
102 pub tables: &'a HashMap<String, BTreeSet<String>>,
103 /// Per fn, the effects whose literal surface is structurally INCOMPLETE — the AS-EFF-008 fail-closed
104 /// marker, without which a benign visible literal masks an invisible forbidden endpoint.
105 pub surface_incomplete: &'a HashMap<String, BTreeSet<E>>,
106 /// Per fn, the TRANSITIVE reason-class tokens — the model's `D` (§6.2 ⟨0.19⟩). The Unknown EFFECT
107 /// propagates along the call graph, so its REASON must too: else `deny E Unknown[reflect]` at a
108 /// caller inheriting Unknown from a reflect-caused callee sees no class and does NOT fire.
109 pub reason_classes: &'a HashMap<String, BTreeSet<String>>,
110 /// Per `Net`-bearing fn, its ⟨0.20⟩ destination classes, ALREADY derived — by [`net_classes_of`] on
111 /// the scan route, read verbatim from the report's `netClass` on the report route. Absent ⇒ empty.
112 pub net_classes: &'a HashMap<String, Vec<String>>,
113}
114
115/// ⟨0.24⟩ ONE `(rule, function)` THE GATE COULD NOT EVALUATE — SPEC §3.1: *"a rule FIRES on a function
116/// only where the match is evidenced by that function's own entry, and is WITHHELD exactly where it is
117/// not. Withholding is per `(rule, function)`, never whole-policy."*
118///
119/// A withheld pair is NOT a tolerated one. Tolerating means the evidence was read and did not match;
120/// withholding means there was no evidence to read, and the two must not arrive at a consumer wearing the
121/// same face. The caller decides the disposition — a violation elsewhere dominates (exit 1, disclose), a
122/// sole withholding is a refusal (exit 2) — but it can only do that if the fact reaches it, which is why
123/// this rides out of [`gate`] beside the violations instead of being logged here.
124#[derive(Debug, Clone, PartialEq, Eq)]
125pub struct Withheld {
126 /// The rule's source line, verbatim (`PolicyRule::raw`).
127 pub rule: String,
128 /// The function the rule could not be evaluated ON. The same rule may fire on another.
129 pub func: String,
130 /// Which narrowing filter had nothing to read — `"Unknown"` or `"Net"`.
131 pub filter: &'static str,
132}
133
134/// ⟨0.24⟩ What [`gate`] returns: the violations it is SURE of, and the `(rule, function)` pairs it
135/// WITHHELD. Both halves travel, because the verdict is both (SPEC §3.1).
136#[derive(Debug, Default)]
137pub struct GateOutcome {
138 /// Sorted by (rule, detail).
139 pub violations: Vec<GateViolation>,
140 /// Sorted by (rule, func). Empty on every policy whose filters the signature can answer.
141 pub withheld: Vec<Withheld>,
142 /// ⟨0.27⟩ SPEC §4 — the RAW TEXT of every rule whose SCOPE bound no function, sorted. A rule that
143 /// bound nothing was evaluated and matched nothing, so it cannot have caught anything; scoring it as
144 /// satisfied makes a one-character typo in a layer name a permanently green gate. This is a
145 /// DISCLOSURE beside the verdict, never a new verdict: the caller prints it and MUST NOT let it
146 /// change the exit code (a zero-match rule is legitimate when one policy is shared across repos).
147 pub zero_match: Vec<String>,
148}
149
150/// ⟨0.24⟩ What one §6.2 `deny`/`pure` rule DOES to one function's signature — see [`rule_hits`].
151pub struct RuleHits<'a> {
152 /// The effects this rule CHARGES on this function, after both narrowing filters. Empty ⇒ the rule
153 /// does not fire here, which is what the disclosure calls PASSING.
154 pub hits: Vec<&'a str>,
155 /// The filters — `"Unknown"` / `"Net"` — that had no evidence to read, in that order. The hit was
156 /// dropped from `hits` AND the fact rides out, because dropping it silently is the mirror defect.
157 pub withheld: Vec<&'static str>,
158}
159
160/// ⟨0.24⟩ WHAT ONE §6.2 `deny`/`pure` RULE DOES TO ONE FUNCTION'S SIGNATURE — the firing decision,
161/// extracted so it has exactly one implementation.
162///
163/// **WHY IT IS A FUNCTION NOW.** It was inline in [`gate`], and the provable-purity disclosure
164/// ([`crate::policy::unverified_hole_rule`]) carried its own second copy that asked a coarser question:
165/// "does the rule NAME an effect this function has?", computed from `r.effects` alone. That copy could
166/// not see a narrowing filter, so it disagreed with the gate on exactly the rules the ⟨0.24⟩ rung added
167/// — and the disagreement ran the LOSING way. A hole is a function that PASSES its rule while `Unknown`,
168/// so a rule the gate TOLERATES (`deny Unknown[reflect]` over an `indirect` hole) was read by the
169/// disclosure as a violation-that-isn't and dropped from the output: `unverified` printed **"every
170/// function in a pure/deny layer is PROVABLY clean ✓"** over a function the gate had just declined to
171/// clear. MEASURED 2026-07-28 on a one-function crate; reachable with NO alias in play, one layer below
172/// the alias-widening defect `ea0df4f` closed.
173///
174/// The caller does the SCOPE test and owns the disposition of `withheld`; this answers only "given that
175/// this rule governs this function, what does it charge, and what could it not evaluate?".
176///
177/// `reason_classes` is the ACCUMULATED (post-fixpoint) class set — `None`/empty means the signature
178/// carries none, which is NOT determinable and is withheld, never floored. `net_classes` is the fn's
179/// ⟨0.20⟩ destination classes, likewise already derived.
180pub fn rule_hits<'a>(
181 r: &PolicyRule,
182 effects: &[&'a str],
183 reason_classes: Option<&BTreeSet<String>>,
184 net_classes: &[String],
185) -> RuleHits<'a> {
186 let mut withheld: Vec<&'static str> = Vec::new();
187 let mut hits: Vec<&str> = if r.effects.is_empty() {
188 // `pure` — every EFFECT, but NOT `Unknown`: the §4 trust marker is not an effect
189 // (AS-EFF-003's concern; `deny Unknown <scope>` is the explicit knob). The reference
190 // engine and the deep backend exclude it identically — this engine wrongly counted an
191 // Unknown-only fn as a `pure` violation until 2026-07-09 (a cross-engine verdict split
192 // on the same policy file).
193 effects.iter().copied().filter(|e| *e != "Unknown").collect()
194 } else {
195 effects.iter().copied().filter(|e| r.effects.contains(e)).collect()
196 };
197 // Reason-scoped Unknown: a `deny E Unknown[classes]` (non-empty filter) keeps its Unknown hit
198 // ONLY for a fn whose TRANSITIVE reason classes include one of those classes; else tolerate it
199 // (wrong reason-class). Concrete effects in `hits` are untouched — only Unknown is scoped.
200 if hits.contains(&"Unknown") && !r.unknown_classes.is_empty() {
201 let want: BTreeSet<&str> = r.unknown_classes.iter().map(|c| c.token()).collect();
202 // An Unknown with NO recorded reason is `unresolved` (conservative — stays in
203 // `[*]`/`[unresolved]`). THIS IS A NET, NOT A ROUTE. It is per FUNCTION and keys on the
204 // ABSENCE of a class set, so any other reason on the same function hides whatever it was
205 // covering — which is how a reasonless chained-dep `Unknown` went ungated on every consumer
206 // that also had a reason of its own. That case now CONTRIBUTES `unresolved` where the
207 // signature is BUILT (candor-scan's `reason_class_direct`; `gate_input_from_report` on the
208 // report route) instead of arriving here by absence. What is left for the absence arm is
209 // the RELEASE-mode gap: the writer's §4 invariant is a `debug_assert`, so a future path
210 // that puts `Unknown` into `direct` with no reason fails closed here rather than escaping
211 // the gate. Not dead — it is pinned by
212 // `reason_scoped_unknown_gate_fires_on_match_tolerates_mismatch`.
213 //
214 // ⟨0.24⟩ The rule itself lives in `crate::policy::reason_class_matches` because
215 // `unverified --class` must select over exactly the set this gate scopes over: a gate and
216 // the disclosure naming the holes that gate did not prove, disagreeing, is the defect.
217 //
218 // ⟨0.24⟩ **BUT THE FLOOR IS ASKED FIRST, AND SEPARATELY — SPEC §3.1.** `reason_class_matches`
219 // answers "could this rule apply?", and its absence/empty arm floors at `unresolved` so a
220 // hole nobody classified never slips out of a filter that names its own class. That is the
221 // right fail-closed default for a MATCHER and the WRONG basis for a FIRING: read as grounds
222 // to emit a violation it asserts a reason NOBODY RECORDED. The two questions shared this one
223 // helper safely only while the report route's refusal short-circuited before `gate()` ran;
224 // `8b97e5c` removed that short-circuit (correctly — a certain violation must reach the
225 // document) and the identical constant, on identical data, became a FABRICATION.
226 //
227 // MEASURED 2026-07-28, `deny Unknown[unresolved] app.opaque` over an entry with `inferred:
228 // ["Unknown"]` and no `direct`, no `unknownWhy`, no `calls`: exit 1 with a violation record
229 // in `--gate-json`, for a function whose determinable class set is EMPTY. The record was
230 // self-refuting — it carried no `reasonClass` key at all, because the floor exists only
231 // inside the predicate and never in the data.
232 //
233 // So the three-way split. NOT determinable ⇒ **WITHHELD**: the hit is dropped AND the pair
234 // rides out to the caller, because dropping it silently is the mirror defect (a narrowed
235 // filter tolerating for lack of evidence is the fail-open this whole rung exists to close).
236 // Determinable ⇒ the shared matcher decides, unchanged, and the `Some(cs)` arm it lands on
237 // is the only one a firing may rest on.
238 //
239 // THE MIRROR IS PINNED, because this is where an under-report gets introduced: an entry
240 // whose `unresolved` is INHERITED — a `calls` edge to a reasonless direct `Unknown` — has a
241 // determinable set of `{unresolved}` (contributed at the ENTRY, before the fixpoint) and
242 // MUST still fire. That is `R1_EXPECT["unresolved"]`'s `app.a_reasonless_only`, and
243 // `a_withheld_unknown_filter_does_not_take_the_inherited_one_with_it` beside it.
244 let classes = reason_classes;
245 let determinable = classes.is_some_and(|cs| !cs.is_empty());
246 if !determinable {
247 hits.retain(|e| *e != "Unknown");
248 withheld.push("Unknown");
249 } else if !reason_class_matches(classes, &want) {
250 hits.retain(|e| *e != "Unknown");
251 }
252 }
253 // Net destination-class: a `deny Net[dest…]` (non-empty filter) keeps its Net hit ONLY for a fn
254 // reaching one of those destination classes; else tolerate (only asserted-safe destinations).
255 // Fail-closed: a masked surface / a Net with no visible host is unknown-host (net_classes_of).
256 //
257 // ⟨0.24⟩ SAME THREE-WAY SPLIT AS THE REASON FILTER, and this side is where the shape is easiest
258 // to see because it never fabricated: with no destination classes to read, `any()` over the
259 // empty set is false and the Net hit was DROPPED — the *other* half of the same defect, an
260 // absence-keyed relaxation of a fail-closed gate. Silently tolerating and silently charging are
261 // the two ways to answer a question the evidence cannot settle; WITHHOLDING is the third, and
262 // the only one that stays true. Costs nothing on a signature this engine produced:
263 // `net_classes_of` floors every Net-bearing fn at `unknown-host`, so an empty set here means
264 // "this producer did not carry the field", never "this function reaches nothing".
265 if hits.contains(&"Net") && !r.net_classes.is_empty() {
266 let fn_net = net_classes;
267 if fn_net.is_empty() {
268 hits.retain(|e| *e != "Net");
269 withheld.push("Net");
270 } else if !fn_net.iter().any(|c| r.net_classes.contains(c)) {
271 hits.retain(|e| *e != "Net");
272 }
273 }
274 RuleHits { hits, withheld }
275}
276
277
278/// Apply a parsed §6.2 policy to an already-accumulated signature. THE ONLY matching code in the stable
279/// toolchain — `candor-scan --policy` and `candor-query gate --report` both land here, which is what
280/// makes "the same verdict from the same signature" a property of the code rather than of two
281/// consistent authors. Returns the violations, sorted by (rule, detail), AND the withheld pairs.
282pub fn gate<E: AsRef<str> + Ord>(p: &ParsedPolicy, gi: &GateInput<E>) -> GateOutcome {
283 let empty: BTreeSet<E> = BTreeSet::new();
284 let no_classes: Vec<String> = Vec::new();
285 let mut out = Vec::new();
286 let mut withheld: Vec<Withheld> = Vec::new();
287 // ONE VERDICT PER (rule, function), whatever the caller's enumeration. `all` is a list of UNITS on
288 // the scan route, and two units can share one qualified name — `#[cfg(unix)] fn f` beside
289 // `#[cfg(not(unix))] fn f` is the everyday case. Their signatures were already merged into one
290 // `inferred` entry keyed by that name, so the gate saw ONE signature and reported it TWICE: two
291 // byte-identical `GateViolation` records, an inflated `N policy violation(s)` count, and a
292 // `--gate-json` document that could not be equal to the one the ⟨0.24⟩ report route produces (a
293 // report is keyed by name, so the duplicate is not reachable there). FOUND BY the §3.1 byte-equality
294 // obligation, on 15 of 90 rows over ebman/pgman/the candor workspace — which is the whole argument
295 // for the verb: no end-to-end test could have separated this from a classifier defect.
296 let mut seen_fn: std::collections::HashSet<&str> = std::collections::HashSet::new();
297 for q in gi.all {
298 // ⟨0.32⟩ the key identifies the unit; the NAME is what a human reads and a scope matches.
299 let disp_q = gi.disp(q);
300 if !seen_fn.insert(q.as_str()) {
301 continue;
302 }
303 let inf = gi.inferred.get(q).unwrap_or(&empty);
304 // Materialized ONCE per function rather than per (function, rule): `rule_hits` is generic over
305 // nothing, so the two routes' effect representations (`&'static str` interned / `String` off the
306 // wire) converge here instead of inside the matcher.
307 let effs: Vec<&str> = inf.iter().map(AsRef::as_ref).collect();
308 // AS-EFF-006 — deny/pure: forbidden effects in the transitive set.
309 for r in &p.rules {
310 if let Some(s) = &r.scope {
311 if !scope_matches(gi.disp(q), s) {
312 continue;
313 }
314 }
315 // ONE implementation of the firing decision, shared with the provable-purity disclosure
316 // ([`crate::policy::unverified_hole_rule`]) — see [`rule_hits`] for what the second copy cost.
317 let RuleHits { hits, withheld: wh } = rule_hits(
318 r,
319 &effs,
320 gi.reason_classes.get(q),
321 gi.net_classes.get(q).map(Vec::as_slice).unwrap_or(&no_classes),
322 );
323 for filter in wh {
324 withheld.push(Withheld { rule: r.raw.clone(), func: gi.disp(q).to_string(), filter });
325 }
326 if !hits.is_empty() {
327 // §6.2: when Unknown is denied, report ALL reason classes on the fn (transitive), so the
328 // consumer sees every reason the strict gate bit — not just the class the rule matched.
329 let reason_class = if hits.contains(&"Unknown") {
330 gi.reason_classes.get(q).map(|cs| cs.iter().cloned().collect()).unwrap_or_default()
331 } else {
332 Vec::new()
333 };
334 // ⟨0.20⟩ when Net is denied, report ALL of the fn's destination classes (transitive).
335 let net_class = if hits.contains(&"Net") {
336 gi.net_classes.get(q).cloned().unwrap_or_default()
337 } else {
338 Vec::new()
339 };
340 out.push(GateViolation {
341 rule: "AS-EFF-006".into(),
342 func: gi.disp(q).to_string(),
343 // ⟨0.32⟩ THE UNIT, beside the name — SPEC §2. Two members of a workspace violating
344 // one rule under one name produced two byte-identical rows until this line.
345 hash: gi.unit(q).to_string(),
346 effects: hits.iter().map(|s| s.to_string()).collect(),
347 detail: format!("`{disp_q}` performs {{ {} }}, forbidden by policy: `{}`", hits.join(", "), r.raw),
348 reason_class,
349 net_class,
350 });
351 }
352 }
353 // AS-EFF-008 — literal allowlists over the transitive literal surfaces.
354 for r in &p.allow_rules {
355 if let Some(s) = &r.scope {
356 if !scope_matches(gi.disp(q), s) {
357 continue;
358 }
359 }
360 if !inf.iter().any(|e| e.as_ref() == r.effect) {
361 continue;
362 }
363 let lits = match r.effect {
364 // `Llm` ⟨0.13⟩ rides the Net host surface (SPEC §1) — `allow Llm <host>` certifies the same
365 // captured hosts as `allow Net`, restricted to the MODEL hosts (a model call's host WAS
366 // captured as a Net literal). Matches candor-java's checkAllowlist("Llm", hostFixpoint, …).
367 "Net" | "Llm" => gi.hosts.get(q),
368 "Exec" => gi.cmds.get(q),
369 "Db" => gi.tables.get(q),
370 _ => gi.paths.get(q),
371 };
372 // An INCOMPLETE surface (a structurally-invisible reach) can't be certified even with visible
373 // hosts — else a benign literal masks the invisible forbidden endpoint (the masking evasion).
374 // `Llm` keys off the NET incompleteness (it rides the Net host literal): a runtime/masked model
375 // host that fails-closes Net must fail-close `allow Llm` too (incompleteAsLlm in candor-java).
376 let inc_key = if r.effect == "Llm" { "Net" } else { r.effect };
377 let surface_incomplete =
378 gi.surface_incomplete.get(q).is_some_and(|s| s.iter().any(|e| e.as_ref() == inc_key));
379 match lits {
380 Some(ls) if !ls.is_empty() && !surface_incomplete => {
381 let bad: Vec<&str> =
382 ls.iter().filter(|l| !literal_allowed(r.effect, l, &r.literals)).map(String::as_str).collect();
383 if !bad.is_empty() {
384 out.push(GateViolation {
385 rule: "AS-EFF-008".into(),
386 func: gi.disp(q).to_string(),
387 hash: gi.unit(q).to_string(), // ⟨0.32⟩ SPEC §2 — every verdict row
388 effects: vec![r.effect.to_string()],
389 detail: format!("`{disp_q}` reaches {{ {} }} outside the allowlist: `{}`", bad.join(", "), r.raw),
390 ..Default::default()
391 });
392 }
393 }
394 _ => out.push(GateViolation {
395 rule: "AS-EFF-008".into(),
396 func: gi.disp(q).to_string(),
397 hash: gi.unit(q).to_string(), // ⟨0.32⟩ SPEC §2 — every verdict row
398 effects: vec![r.effect.to_string()],
399 detail: format!("`{disp_q}` performs {} with no visible literal — the surface cannot be certified: `{}`", r.effect, r.raw),
400 ..Default::default()
401 }),
402 }
403 }
404 // AS-EFF-009 — layering: no fn in scope A may transitively reach scope B.
405 for r in &p.layer_rules {
406 if !scope_matches(gi.disp(q), &r.from) {
407 continue;
408 }
409 let mut seen: BTreeSet<&str> = BTreeSet::new();
410 let mut stack: Vec<&str> =
411 gi.calls.get(q).map(|cs| cs.iter().map(String::as_str).collect()).unwrap_or_default();
412 let mut hit: Option<&str> = None;
413 while let Some(n) = stack.pop() {
414 if !seen.insert(n) {
415 continue;
416 }
417 if scope_matches(gi.disp(n), &r.to) {
418 hit = Some(n);
419 break;
420 }
421 if let Some(cs) = gi.calls.get(n) {
422 stack.extend(cs.iter().map(String::as_str));
423 }
424 }
425 if let Some(h) = hit {
426 out.push(GateViolation {
427 rule: "AS-EFF-009".into(),
428 func: gi.disp(q).to_string(),
429 hash: gi.unit(q).to_string(), // ⟨0.32⟩ SPEC §2 — every verdict row
430 effects: Vec::new(), // a layer-flow has no single effect
431 detail: format!("`{disp_q}` reaches into a forbidden layer (via `{h}`): `{}`", r.raw),
432 ..Default::default()
433 });
434 }
435 }
436 // ⟨0.29⟩ AS-EFF-011 — `only A -> B …`: a fn in A may reach A and the listed scopes, NOTHING else.
437 //
438 // The same walk as `forbid` above with the test INVERTED, and the inversion is the point rather
439 // than the code. `forbid` fails OPEN — what you did not prohibit is permitted — so a leaf package
440 // can only be protected by enumerating what it must not reach, and that list does not cover a
441 // package added tomorrow. `only` fails SAFE: the dependency you forgot to permit is a violation,
442 // loudly, on the day it appears.
443 //
444 // THE WALK STOPS AT A PERMITTED SCOPE. A permitted callee's own dependencies are governed by the
445 // rules about IT; descending past it would make `only` demand the transitive closure of everything
446 // you permit, which is the same enumeration-that-rots one level down. `from` IS descended through
447 // — a fn in A calling another fn in A that reaches infra is still A reaching infra.
448 for r in &p.only_rules {
449 if !scope_matches(gi.disp(q), &r.from) {
450 continue;
451 }
452 let mut seen: BTreeSet<&str> = BTreeSet::new();
453 let mut stack: Vec<&str> =
454 gi.calls.get(q).map(|cs| cs.iter().map(String::as_str).collect()).unwrap_or_default();
455 let mut hit: Option<&str> = None;
456 while let Some(n) = stack.pop() {
457 if !seen.insert(n) {
458 continue;
459 }
460 // ⟨0.29⟩ EXACT segment match on a PERMITTED scope — see `scope_matches_permitted`. The
461 // shared prefix matcher is fail-CLOSED for every other rule kind and fail-OPEN here.
462 if r.to.iter().any(|t| scope_matches_permitted(gi.disp(n), t)) {
463 continue; // permitted, and its own callees are not this rule's business
464 }
465 if !scope_matches(gi.disp(n), &r.from) {
466 hit = Some(n);
467 break;
468 }
469 if let Some(cs) = gi.calls.get(n) {
470 stack.extend(cs.iter().map(String::as_str));
471 }
472 }
473 if let Some(h) = hit {
474 out.push(GateViolation {
475 // ⟨0.29⟩ ITS OWN CODE, not `forbid`'s. A rule code is the handle a CI suppression, a
476 // dashboard link and an alert filter key on, and these two are opposite constructs —
477 // must-not-reach versus must-be-on-the-list. Sharing 009 would make every existing
478 // `forbid` suppression silently start muting `only` violations its author never
479 // accepted: a fail-open change to an operator's config, made by us and invisible to
480 // them, which is the argument this form is built on turned on the tool.
481 rule: "AS-EFF-011".into(),
482 func: gi.disp(q).to_string(),
483 hash: gi.unit(q).to_string(), // ⟨0.32⟩ SPEC §2 — every verdict row
484 effects: Vec::new(),
485 detail: format!(
486 "`{disp_q}` reaches `{h}`, which this permission rule does not permit: `{}`",
487 r.raw
488 ),
489 ..Default::default()
490 });
491 }
492 }
493 }
494 // Sort by (rule, detail) — identical order to the old rendered-line sort (the "[rule] detail" render
495 // puts the constant '[' first and all AS-EFF codes are same-length), without allocating two Strings
496 // per comparison.
497 //
498 // ⟨0.32⟩ …AND THEN BY UNIT (SPEC §2). Two units sharing a name produce rows whose `rule` and
499 // `detail` are identical — `detail` is built from the DISPLAY name — so the pair tied and their
500 // order was whatever `all` happened to hold. §3.3.1 makes the order part of the byte-equality
501 // between the two routes, and the two routes accumulate in different orders. The same key is
502 // applied again in `candor_report`'s verdict writers, where the SCAN route's cross-member
503 // concatenation is re-sorted.
504 out.sort_by(|a, b| {
505 (a.rule.as_str(), a.detail.as_str(), a.hash.as_str())
506 .cmp(&(b.rule.as_str(), b.detail.as_str(), b.hash.as_str()))
507 });
508 // Deterministic for the same reason the violations are: a disclosure a consumer diffs between runs
509 // must not reorder because a HashMap iterated differently.
510 withheld.sort_by(|a, b| (&a.rule, &a.func).cmp(&(&b.rule, &b.func)));
511 withheld.dedup();
512 // ⟨0.27⟩ ZERO-MATCH DISCLOSURE. Counted over the SAME key set the gate iterated, so "bound nothing"
513 // means here exactly what it means to the gate. A `deny`/`pure` with NO scope applies to every
514 // function and so can never be this kind of typo — excluded. A layer rule counts a match on either
515 // endpoint, over the call-graph keys it binds across.
516 let mut zero: std::collections::BTreeMap<&str, usize> = std::collections::BTreeMap::new();
517 for r in &p.rules {
518 if r.scope.is_some() {
519 zero.entry(r.raw.as_str()).or_insert(0);
520 }
521 }
522 for r in &p.layer_rules {
523 zero.entry(r.raw.as_str()).or_insert(0);
524 }
525 // ⟨0.29⟩ an `only` rule binds nothing when NEITHER endpoint names anything in this tree — the same
526 // typo channel `forbid` has, and the more dangerous one to leave silent: a `forbid` that binds
527 // nothing merely fails to prohibit, while an `only` that binds nothing withholds a promise the
528 // operator believes they made.
529 for r in &p.only_rules {
530 zero.entry(r.raw.as_str()).or_insert(0);
531 }
532 if !zero.is_empty() {
533 let mut names: std::collections::BTreeSet<&str> =
534 gi.all.iter().map(|q| q.as_str()).collect();
535 names.extend(gi.calls.keys().map(String::as_str));
536 for n in names {
537 for r in &p.rules {
538 if let Some(s) = &r.scope {
539 if scope_matches(gi.disp(n), s) {
540 *zero.entry(r.raw.as_str()).or_insert(0) += 1;
541 }
542 }
543 }
544 for r in &p.layer_rules {
545 if scope_matches(gi.disp(n), &r.from) || scope_matches(gi.disp(n), &r.to) {
546 *zero.entry(r.raw.as_str()).or_insert(0) += 1;
547 }
548 }
549 // ON `from` ONLY, and deliberately NOT on either endpoint the way a `forbid` counts. A
550 // `forbid`'s subject is the pair; an `only`'s subject is `from` — it is a promise ABOUT that
551 // scope — so a rule whose destinations all exist while its `from` names nothing has bound
552 // nothing at all, and is exactly the typo that leaves an operator believing a leaf is
553 // protected. Counting the destinations would hide it behind a scope that happens to resolve.
554 for r in &p.only_rules {
555 if scope_matches(gi.disp(n), &r.from) {
556 *zero.entry(r.raw.as_str()).or_insert(0) += 1;
557 }
558 }
559 }
560 }
561 let zero_match: Vec<String> = zero
562 .into_iter()
563 .filter(|(_, c)| *c == 0)
564 .map(|(raw, _)| raw.to_string())
565 .collect();
566 GateOutcome { violations: out, withheld, zero_match }
567}