@clear-capabilities/agentic-security-scanner 0.136.2 → 0.137.0

package/src/posture/calibration-drift.js

@@ -57,12 +57,18 @@ export function computeDrift(scanRoot, opts = {}) {
   const fb = loadTriageFeedback(scanRoot);
   if (!fb.length) return { alarms: [], note: 'no-feedback-data' };
 
-  // Group by family. Each entry should carry: family, verdict ('tp'|'fp'|'wai'),
-  // reportedConfidence (0..1), ts.
+  // Group by family. Each entry should carry: family, verdict ('tp'|'fp'|'wontfix'),
+  // reportedConfidence (0..1), at.
+  // Field names/verdict values here previously didn't match the real
+  // producer (commands/triage.md's Step 2): it writes `at`, not `ts`, and
+  // never wrote `reportedConfidence` at all (fixed there alongside this —
+  // Stage 2 measurement-completeness audit) or a `'wai'` verdict (it only
+  // ever writes 'tp'|'fp'|'wontfix'), so this alarm could never fire against
+  // real triage data regardless of how badly calibration had drifted.
   const byFamily = new Map();
   for (const e of fb) {
-    if (!e || !e.family || !inWindow(e.ts, window)) continue;
-    if (!['tp', 'fp', 'wai'].includes(e.verdict)) continue;
+    if (!e || !e.family || !inWindow(e.at, window)) continue;
+    if (!['tp', 'fp', 'wontfix'].includes(e.verdict)) continue;
     if (typeof e.reportedConfidence !== 'number') continue;
     if (!byFamily.has(e.family)) byFamily.set(e.family, []);
     byFamily.get(e.family).push(e);
@@ -75,7 +81,7 @@ export function computeDrift(scanRoot, opts = {}) {
     const reportedAcc = entries.reduce((acc, e) => acc + e.reportedConfidence, 0) / entries.length;
     const divergence = Math.abs(reportedAcc - realizedAcc);
     if (divergence < threshold) continue;
-    const firstTs = entries.map(e => e.ts).filter(Boolean).sort()[0];
+    const firstTs = entries.map(e => e.at).filter(Boolean).sort()[0];
     alarms.push({
       alarm: true,
       since: firstTs,

package/src/posture/calibration.js

@@ -121,15 +121,37 @@ export function loadCalibrationHistory(scanRoot) {
   };
   if (seed) merge(seed);
   if (customer) merge(customer);
-  // Merge triage-derived TP/FP counts (auto-feedback loop)
+  // Merge triage-derived TP/FP counts (auto-feedback loop).
+  //
+  // EA-01: only a state a human actually chose is a real signal. triage.js's
+  // syncWithScan sets every new finding to 'open' with no human involved, and
+  // auto-closes any finding the scanner stops seeing to 'fixed' with
+  // `automatic:true` on the transition — neither reflects a judgment that the
+  // finding was real. Counting them as TP made the "true positive" column
+  // entirely machine-generated while the FP column required an explicit human
+  // false-positive mark, so calibrated confidence trended toward 1.0 for any
+  // family with enough scan volume regardless of actual accuracy. Only a
+  // manual 'fixed' transition (confirmed real, now resolved) counts as TP;
+  // only 'false-positive' counts as FP. 'open', 'in-progress', 'wont-fix',
+  // and an automatically-reached 'fixed' contribute to neither bucket.
   try {
     const triage = _readJsonMaybe(statePath(scanRoot, 'triage.json'));
     if (triage && triage.findings) {
+      const lastFixedIsAutomatic = new Map();
+      for (const t of triage.transitions || []) {
+        if (t && t.to === 'fixed') lastFixedIsAutomatic.set(t.id, t.automatic === true);
+      }
       const triageFams = {};
       for (const f of Object.values(triage.findings)) {
         const fam = f.family || 'unknown';
         if (!triageFams[fam]) triageFams[fam] = { tp: 0, fp: 0 };
-        if (f.state === 'fixed' || f.state === 'open' || f.state === 'in-progress') triageFams[fam].tp++;
+        // Require POSITIVE proof of a manual transition (=== false), not
+        // merely the absence of proof of an automatic one (!== true) — the
+        // latter fails OPEN: a finding with no transition record at all
+        // (lastFixedIsAutomatic.get returns undefined) counted as a manual
+        // TP by default, exactly the machine-generated-signal failure mode
+        // this whole block's comment states it exists to prevent.
+        if (f.state === 'fixed' && lastFixedIsAutomatic.get(f.id) === false) triageFams[fam].tp++;
         else if (f.state === 'false-positive') triageFams[fam].fp++;
       }
       merge({ families: triageFams });

package/src/posture/clustering.js

@@ -28,7 +28,18 @@ function sinkKey(f) {
   const parser = f.parser || '';
   const rule = f.cwe || f.family || (f.vuln || '').slice(0, 40);
   const file = f.file || f.sink?.file || '';
-  const sinkExpr = (f.sink?.label || f.sink?.snippet || f.snippet || '')
+  // PRD R3: dataflow/engine.js sets sink.label to the catalog rule id
+  // (f.sinkId) for IR-TAINT findings — the same generic string for every
+  // callsite of that rule, not a per-callsite snippet. Two textually-unrelated
+  // sinks sharing a rule (e.g. two independent eval() calls in one file) would
+  // otherwise collapse into one cluster and silently drop a real finding. The
+  // dedup pass upstream (dedupeFindingsWithEvidence) already collapses
+  // multiple sources converging on the SAME sink line, so every IR-TAINT
+  // finding reaching this point has a distinct line — folding the line into
+  // the key here can only split buckets more finely, never wrongly merge one.
+  const sinkExpr = (parser === 'IR-TAINT'
+    ? `L${f.sink?.line || f.line || 0}:${f.sink?.label || ''}`
+    : (f.sink?.label || f.sink?.snippet || f.snippet || ''))
     .replace(/['"`][^'"`]*['"`]/g, '_S_')
     .replace(/\s+/g, ' ')
     .trim()

package/src/posture/compliance-frameworks/nist-privacy-1-1.json

@@ -781,7 +781,7 @@
       "summary": "Software is maintained, replaced, and removed commensurate with risk.",
       "codeTestable": "partial",
       "mapsTo": [
-        "family:vulnerable-dependency",
+        "family:vulnerable-dep",
         "family:dependency-drift"
       ]
     },
@@ -800,7 +800,7 @@
       "codeTestable": "yes",
       "mapsTo": [
         "family:dependency-confusion",
-        "family:vulnerable-dependency"
+        "family:vulnerable-dep"
       ]
     },
     {

package/src/posture/compliance-frameworks/owasp-asvs-5.json

@@ -66,7 +66,7 @@
       "category": "Malicious Code",
       "summary": "Verify the application does not include known-malicious or compromised dependencies.",
       "evidence": [".agentic-security/sbom-history snapshots clean.", "No dependency-confusion or dependency-drift findings."],
-      "mapsTo": ["family:vulnerable-dependency", "family:dependency-confusion", "family:dependency-drift"]
+      "mapsTo": ["family:vulnerable-dep", "family:dependency-confusion", "family:dependency-drift"]
     },
     {
       "id": "V14.1",

package/src/posture/compliance-policy.js

@@ -41,6 +41,7 @@ import * as fs from 'node:fs';
 import * as path from 'node:path';
 import * as yaml from '../util/yaml.js';
 import { statePath, safeWriteState } from './state-dir.js';
+import { SCANNER_VERSION } from './version.js';
 
 const POLICY_FILE = 'compliance.policy.yml';
 
@@ -128,7 +129,29 @@ function _runCheck(check, ctx) {
  * Run all controls in the policy and emit a verification report.
  */
 export function verifyPolicy(policy, ctx) {
+  // CMP-5: loadPolicy() reports a parse failure as { _error }, which has no
+  // `.controls` — the same shape as "no policy file at all". Distinguishing
+  // them matters: a customer who typo'd their YAML deserves a loud error,
+  // not a report that silently treats their (unparsed) policy as having no
+  // controls to check.
+  if (policy && policy._error) return { controls: [], status: 'error', error: policy._error };
   if (!policy || !policy.controls) return { controls: [], status: 'no-policy' };
+  // CMP-5: a finding-family check must see every channel a real scan
+  // produces (findings=SAST, secrets, logicVulns, supplyChain=SCA) — the
+  // caller (engine.js) hands last-scan.json's own channel split through
+  // ctx, so a control checking family:hardcoded-secret or
+  // family:vulnerable-dep isn't blind to 3 of the engine's 4 finding types.
+  // Family defaults mirror report/index.js's normalizeFindings so the two
+  // modules can't drift apart on what family an untagged finding belongs to.
+  ctx = {
+    ...ctx,
+    findings: [
+      ...(ctx.findings || []),
+      ...(ctx.secrets || []).map(s => ({ ...s, family: s.family || 'hardcoded-secret' })),
+      ...(ctx.logicVulns || []),
+      ...(ctx.supplyChain || []).map(sc => ({ ...sc, family: sc.family || 'vulnerable-dep' })),
+    ],
+  };
   const results = [];
   for (const control of policy.controls) {
     if (control.not_applicable) {
@@ -166,6 +189,13 @@ export function emitEvidenceJsonLd(report, scanRoot) {
     framework: report.framework,
     version: report.version,
     generatedAt: new Date().toISOString(),
+    // CMP-5: this artifact is fed to GRC tooling (Vanta/Drata/SecureFrame) and
+    // auditors largely unread by a human — the same reason
+    // auditor-walkthrough.js's narrative carries this disclaimer verbatim.
+    // Without it here, a machine-consumed "ComplianceEvidence" document reads
+    // as an attestation, not a scanner's automated observation.
+    disclaimer: 'This artifact organizes automated scanner evidence into a structured report. It does not certify compliance. A licensed assessor (CPA / auditor / DPO) is responsible for the final attestation.',
+    provenance: { engineVersion: SCANNER_VERSION },
     summary: report.summary,
     controls: report.controls.map(c => ({
       '@type': 'Control',
@@ -192,7 +222,9 @@ export function emitEvidenceMarkdown(report, scanRoot) {
   const lines = [];
   lines.push(`# Compliance evidence — ${report.framework}`);
   lines.push('');
-  lines.push(`Generated by agentic-security on ${new Date().toISOString().slice(0,10)}.`);
+  lines.push(`Generated by agentic-security (engine ${SCANNER_VERSION}) on ${new Date().toISOString().slice(0,10)}.`);
+  lines.push('');
+  lines.push('> This document organizes automated scanner evidence into a structured report. It does not certify compliance. A licensed assessor (CPA / auditor / DPO) is responsible for the final attestation.');
   lines.push('');
   lines.push(`Compliant: **${report.summary.compliant}** / Non-compliant: **${report.summary.nonCompliant}** / Not applicable: **${report.summary.notApplicable}** of ${report.summary.total} controls.`);
   lines.push('');

package/src/posture/confidence.js

@@ -47,19 +47,53 @@ export function annotateConfidence(findings) {
       if (f.routeRooted) conf = Math.min(1, conf + 0.05);
       if (f.guards && f.guards.length) conf *= 0.80;
       if (f.reachable === false) conf *= 0.55;
+      // f.unvalidated is set later in the pipeline (llm-validator/index.js,
+      // invoked well after this first confidence pass), so it's never true
+      // here on a finding's first computation — see applyUnvalidatedPenalty
+      // below, the real enforcement point, run after validation. Kept here
+      // too so a caller that builds a synthetic finding with
+      // unvalidated:true pre-set still gets it applied in one pass.
       if (f.unvalidated) conf *= 0.85;   // LLM validator unavailable
-      if (f.llmOnly) conf *= 0.70;       // LLM-only finding, no Layer-2 path
+      // f.llmOnly: currently unreachable — no producer anywhere in this
+      // codebase ever sets it (grepped). Documented as a no-op rather than
+      // silently deleted, since the intent ("LLM-only finding, no Layer-2
+      // path") is a real, plausible signal that just isn't wired yet.
+      if (f.llmOnly) conf *= 0.70;
     }
     conf = Math.max(0, Math.min(1, conf));
     f.confidence = Math.round(conf * 1000) / 1000;
-    // Premortem 3R-15: derive tier from the 2-decimal display value so a
-    // finding reported as "0.75" never lands in two tiers depending on the
-    // viewer's rounding. Add a +0.005 epsilon to anchor cutoffs to the
-    // displayed rounded value (3-decimal raw 0.745 → 2-decimal 0.75 → high).
-    const display = Math.round(f.confidence * 100) / 100;
-    if (display >= 0.75) f.confidenceTier = 'high';
-    else if (display >= 0.50) f.confidenceTier = 'medium';
-    else if (display >= 0.25) f.confidenceTier = 'low';
-    else f.confidenceTier = 'very-low';
+    f.confidenceTier = _tierFor(f.confidence);
+  }
+}
+
+// Premortem 3R-15: derive tier from the 2-decimal display value so a
+// finding reported as "0.75" never lands in two tiers depending on the
+// viewer's rounding (3-decimal raw 0.745 → 2-decimal 0.75 → high).
+function _tierFor(confidence) {
+  const display = Math.round(confidence * 100) / 100;
+  if (display >= 0.75) return 'high';
+  if (display >= 0.50) return 'medium';
+  if (display >= 0.25) return 'low';
+  return 'very-low';
+}
+
+// Retroactively applies the unvalidated penalty to findings whose
+// confidence was already computed by annotateConfidence BEFORE
+// f.unvalidated could be known — annotateConfidence only computes from
+// scratch when f.confidence is still null (so hand-tuned detector
+// confidences survive untouched), and f.unvalidated is set by
+// llm-validator/index.js, which runs well after the pipeline's first
+// annotateConfidence pass. Call this once, immediately after LLM
+// validation runs (or is skipped). Idempotent: a finding is only
+// adjusted once, tracked via f._unvalidatedPenaltyApplied.
+export function applyUnvalidatedPenalty(findings) {
+  if (!Array.isArray(findings)) return;
+  for (const f of findings) {
+    if (!f || typeof f !== 'object') continue;
+    if (f._unvalidatedPenaltyApplied) continue;
+    f._unvalidatedPenaltyApplied = true;
+    if (!f.unvalidated || typeof f.confidence !== 'number') continue;
+    f.confidence = Math.round(Math.max(0, Math.min(1, f.confidence * 0.85)) * 1000) / 1000;
+    f.confidenceTier = _tierFor(f.confidence);
   }
 }

package/src/posture/corpus-enroll.js

@@ -296,8 +296,12 @@ function _stripState(dir) {
   }
 }
 
-// `scoreCandidate` is deliberately NOT exported: an external caller could
-// score a candidate and then write it by some other route, which is exactly
-// the unscored-write path this module exists to make unavailable. Enrolment
-// scores and writes as one operation or not at all.
-export const _internals = { DEFAULT_TIER, scoreCandidate, _languageOf, _escapeRegex, _stripState };
+// `scoreCandidate` is deliberately NOT exported — not even via _internals,
+// the test-only-helpers convention used elsewhere in this codebase: an
+// external caller could score a candidate and then write it by some other
+// route, which is exactly the unscored-write path this module exists to
+// make unavailable. Enrolment scores and writes as one operation or not at
+// all. Keep it out of this object even though every other _internals
+// export in this codebase is a harmless pure helper — this one specific
+// function is the security boundary, not test plumbing.
+export const _internals = { DEFAULT_TIER, _languageOf, _escapeRegex, _stripState };

package/src/posture/corpus-match.js

@@ -54,6 +54,25 @@ function _matches(f, manifest, matcher) {
     (manifest?.cwe ? f.cwe === manifest.cwe || matcher.test(f.cwe || '') : true);
 }
 
+/**
+ * Every finding, across every channel, that scores this entry.
+ *
+ * Third caller of the ONE predicate (gate, enrollment, and now the per-layer
+ * recall instrument in `bench/layer-recall/`). That instrument needs to know
+ * WHICH findings matched so it can attribute them to the analysis layer that
+ * produced them; re-deriving the predicate there is precisely the drift this
+ * module was extracted to prevent, so it reuses `_matches` instead.
+ */
+export function matchingFindings(scan, manifest, matcher = matcherFor(manifest)) {
+  const out = [];
+  for (const channel of CHANNELS) {
+    const arr = scan?.[channel];
+    if (!Array.isArray(arr)) continue;
+    for (const f of arr) if (_matches(f, manifest, matcher)) out.push(f);
+  }
+  return out;
+}
+
 /** Did the vulnerable (`pre/`) tree produce a matching finding? */
 export function preHit(scan, manifest, matcher = matcherFor(manifest)) {
   return _any(scan, f => _matches(f, manifest, matcher));

package/src/posture/csharp-analysis.js

@@ -101,6 +101,49 @@ function isSanitizedExpr(text) {
   return false;
 }
 
+// For each SANITIZER_PATTERNS match in `text`, find the nearest `(...)` call
+// span after the match and return its [start, end) bounds. Used to tell
+// "the sanitizer call actually wraps the tainted value" (HtmlEncode(x)) apart
+// from "the sanitizer pattern matched something unrelated elsewhere in a
+// compound expression" (comment + (IsNullOrEmpty(flag) ? ... : ...)) — a
+// common .NET idiom that combines a tainted value with an unrelated validity
+// check in the same expression.
+function _sanitizerCallSpans(text) {
+  const spans = [];
+  for (const re of SANITIZER_PATTERNS) {
+    const g = new RegExp(re.source, re.flags.includes('g') ? re.flags : re.flags + 'g');
+    let m;
+    while ((m = g.exec(text))) {
+      const openIdx = text.indexOf('(', g.lastIndex - 1);
+      if (openIdx !== -1 && openIdx - g.lastIndex < 3) {
+        let depth = 1, j = openIdx + 1;
+        while (j < text.length && depth > 0) {
+          if (text[j] === '(') depth++;
+          else if (text[j] === ')') depth--;
+          j++;
+        }
+        spans.push([openIdx, j]);
+      }
+      if (g.lastIndex === m.index) g.lastIndex++; // avoid infinite loop on zero-width matches
+    }
+  }
+  return spans;
+}
+
+// Is `ref` tainted-and-NOT-neutralized by an enclosing sanitizer call in
+// `text`? True when every occurrence of `ref` as a whole word sits outside
+// every sanitizer call span (i.e. no sanitizer actually wraps it).
+function _refEscapesSanitizers(text, ref, spans) {
+  if (!spans.length) return true;
+  const re = new RegExp(`\\b${ref.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}\\b`, 'g');
+  let m, sawAny = false;
+  while ((m = re.exec(text))) {
+    sawAny = true;
+    if (!spans.some(([s, e]) => m.index >= s && m.index < e)) return true;
+  }
+  return !sawAny;
+}
+
 // Walk a single method's body and compute per-variable type + taint.
 // Returns { typeMap, taintMap, sourceLines } where sourceLines records the
 // declaration line at which each variable first became tainted.
@@ -296,12 +339,22 @@ export function receiverIsType(method, flow, receiver, typePattern) {
 // for short expressions but unsafe for arbitrary string-containing text.
 export function expressionIsTainted(flow, text, idents = null) {
   if (!text && !idents) return false;
+  // Check known-tainted variable references that ESCAPE every sanitizer
+  // call span first — a whole-expression sanitizer-pattern match (below)
+  // must not clear a tainted variable the sanitizer doesn't actually wrap.
+  // `HtmlEncode(x)` genuinely neutralizes `x` (inside the call's parens);
+  // `comment + (string.IsNullOrEmpty(flag) ? ... : ...)` does not neutralize
+  // `comment` just because an unrelated sanitizer pattern matched `flag`
+  // elsewhere in the same compound expression.
+  const refs = idents || (text ? text.match(/\b[A-Za-z_]\w*\b/g) || [] : []);
+  const spans = text ? _sanitizerCallSpans(text) : [];
+  for (const r of refs) {
+    if (flow.taintMap.get(r) && (!text || _refEscapesSanitizers(text, r, spans))) return true;
+  }
   if (text) {
     if (isSourceExpr(text) && !isSanitizedExpr(text)) return true;
     if (isSanitizedExpr(text)) return false;
   }
-  const refs = idents || (text ? text.match(/\b[A-Za-z_]\w*\b/g) || [] : []);
-  for (const r of refs) if (flow.taintMap.get(r)) return true;
   return false;
 }
 
@@ -311,9 +364,15 @@ export function expressionIsTainted(flow, text, idents = null) {
 // not treated as code identifiers.
 export function argIsTainted(flow, arg) {
   if (!arg) return false;
+  // Same span-aware fix as expressionIsTainted: a tainted identifier that
+  // escapes every sanitizer call span wins over a whole-argument
+  // sanitizer-pattern match.
+  const spans = arg.text ? _sanitizerCallSpans(arg.text) : [];
+  for (const id of arg.idents || []) {
+    if (flow.taintMap.get(id) && (!arg.text || _refEscapesSanitizers(arg.text, id, spans))) return true;
+  }
   if (arg.text && isSanitizedExpr(arg.text)) return false;
   if (arg.text && isSourceExpr(arg.text)) return true;
-  for (const id of arg.idents || []) if (flow.taintMap.get(id)) return true;
   return false;
 }
 

package/src/posture/deploy-platform.js

@@ -4,7 +4,10 @@
 // returns platform-specific security findings: missing headers, public previews,
 // no health checks, unsafe infra settings.
 //
-// Platforms: Vercel, Railway, Fly.io, Render, Netlify, AWS Amplify, Cloudflare
+// Platforms: Vercel, Railway, Fly.io, Netlify, Cloudflare.
+// NOT implemented despite being previously listed here: Render, AWS Amplify
+// (found via Stage-0 doc audit, 2026 — grepped for any render/amplify config
+// path or platform key in this file; zero hits for either).
 
 import * as fs from 'node:fs';
 import * as path from 'node:path';

package/src/posture/drift.js

@@ -8,7 +8,13 @@
 
 function _routeKey(r) { return `${r.method || 'ANY'} ${r.path || '(file)'} @ ${r.file}:${r.line}`; }
 function _depKey(c)   { return `${c.ecosystem}:${c.name}@${c.version}`; }
-function _findingKey(f) { return `${f.kind}:${f.file}:${f.line}:${(f.vuln||'').slice(0,80)}`; }
+// Prefer stableId (posture/stable-id.js) — it omits the exact source line by
+// design, so an unrelated edit that shifts a still-unfixed finding's line
+// number doesn't register as one "removed" + one "added" finding, inflating
+// drift tier and falsely flagging a PR as introducing/fixing something it
+// didn't touch. Falls back to the line-sensitive key only for finding shapes
+// that never got a stableId annotated.
+function _findingKey(f) { return f.stableId || `${f.kind}:${f.file}:${f.line}:${(f.vuln||'').slice(0,80)}`; }
 
 function _toMap(arr, keyFn) {
   const m = new Map();

package/src/posture/epss.js

@@ -6,10 +6,19 @@
 // from those attackers are actively weaponizing.
 //
 // Decoration shape (added to each SCA finding with a CVE):
-//   epss: 0.92345
+//   epssScore: 0.92345   (this comment previously said `epss:`; the code has
+//                         always written `epssScore` — the comment was wrong,
+//                         not the code)
 //   epssPercentile: 0.987
 //   exploitedNow: true   ← percentile >= 0.95
 //
+// ⚠ engine.js ALSO sets f.epssScore/f.epssPercentile directly (around line
+// 6451), independently of this module's fetchEPSS/decorate path — two
+// implementations of the same decoration. This module's own `fetchEPSS` has
+// no in-tree caller (allowlisted in no-dead-modules.test.js as future-public
+// API); the live path is the one in engine.js. Worth consolidating, not
+// resolved here.
+//
 // Source: https://api.first.org/data/v1/epss?cve=CVE-...,CVE-...
 // Cached on disk: ~/.claude/agentic-security/epss-cache/<sha256>.json
 // 24-hour TTL. Falls back gracefully when offline.
@@ -93,6 +102,9 @@ function cvesIn(finding) {
   const found = new Set();
   if (typeof finding.cve === 'string') found.add(finding.cve.toUpperCase());
   if (Array.isArray(finding.cves)) for (const c of finding.cves) found.add(String(c).toUpperCase());
+  // The actual field name every SCA finding in this codebase uses (src/sca/
+  // CLAUDE.md's documented shape; carried through by normalizeFindings).
+  if (Array.isArray(finding.cveAliases)) for (const c of finding.cveAliases) found.add(String(c).toUpperCase());
   if (Array.isArray(finding.vulnerabilities)) {
     for (const v of finding.vulnerabilities) {
       if (typeof v.id === 'string' && v.id.startsWith('CVE-')) found.add(v.id.toUpperCase());

package/src/posture/evidence-bundle.js

@@ -104,12 +104,25 @@ export function ensureKeyPair(dir = keyDir()) {
     return { privateKeyPem, publicKeyPem, created: true, ...p };
   } catch (e) {
     if (e.code !== 'EEXIST') throw e;
-    return {
-      privateKeyPem: fs.readFileSync(p.privateKey, 'utf8'),
-      publicKeyPem: fs.readFileSync(p.publicKey, 'utf8'),
-      created: false,
-      ...p,
-    };
+    // We lost the race on the private key — someone else's 'wx' won.
+    // Their public-key write follows immediately after their private-key
+    // write but is not itself atomic with it, so it may not have landed
+    // yet: a bare read here would throw an uncaught ENOENT on a genuinely
+    // transient state, not a real error. Retry briefly rather than crash.
+    const sleepBuf = new Int32Array(new SharedArrayBuffer(4));
+    for (let attempt = 0; ; attempt++) {
+      try {
+        return {
+          privateKeyPem: fs.readFileSync(p.privateKey, 'utf8'),
+          publicKeyPem: fs.readFileSync(p.publicKey, 'utf8'),
+          created: false,
+          ...p,
+        };
+      } catch (readErr) {
+        if (readErr.code !== 'ENOENT' || attempt >= 50) throw readErr;
+        Atomics.wait(sleepBuf, 0, 0, 10); // 10ms; ~500ms total budget
+      }
+    }
   }
 }
 
@@ -227,9 +240,26 @@ export function signEvidenceBundle(bundle, privateKeyPem) {
  * Returns {ok, reason}. Never throws — a malformed bundle from an untrusted
  * source is an expected input, not an exceptional one.
  */
+// The complete set of top-level keys a legitimately-built, signed bundle can
+// carry — buildEvidenceBundle's six plus signEvidenceBundle's `signature`.
+// EA-03 (Stage-0 audit, 2026): canonicalBytes SIGNS an allowlist of fields;
+// verifyEvidenceBundle never checked for keys OUTSIDE that allowlist, so a
+// bundle with a fabricated `verdict`/`proofLevel`/anything-else stapled on
+// after signing verified as authentic — the signature simply never covered
+// those bytes. Rejecting unknown keys here closes that; it must exactly match
+// what buildEvidenceBundle+signEvidenceBundle actually produce, or a
+// legitimate bundle would start failing verification.
+const BUNDLE_TOP_LEVEL_KEYS = new Set([
+  'schema', 'finding', 'evidence', 'engine', 'proves', 'doesNotProve', 'signature',
+]);
+
 export function verifyEvidenceBundle(bundle, publicKeyPem) {
   if (!bundle || typeof bundle !== 'object') return { ok: false, reason: 'bundle is not an object' };
   if (bundle.schema !== BUNDLE_SCHEMA) return { ok: false, reason: `unrecognised schema: ${bundle.schema}` };
+  const unknownKeys = Object.keys(bundle).filter(k => !BUNDLE_TOP_LEVEL_KEYS.has(k));
+  if (unknownKeys.length) {
+    return { ok: false, reason: `unrecognised top-level key(s) not covered by the signature: ${unknownKeys.join(', ')}` };
+  }
   const sig = bundle.signature;
   if (!sig?.value) return { ok: false, reason: 'bundle is unsigned' };
   if (sig.algorithm !== 'ed25519') return { ok: false, reason: `unsupported algorithm: ${sig.algorithm}` };

package/src/posture/exploitability-probability.js

@@ -189,9 +189,21 @@ export function annotateExploitProbability(findings, ctx = {}) {
     p = _clamp01(p);
     // Wilson CI: prefer historical CI when we have one, otherwise derive
     // a wider CI from the prior (n=10 implied sample at base rate).
+    //
+    // The point estimate (exploitProbability) is ALWAYS the per-finding,
+    // factor-adjusted `p` computed above — never the population-level
+    // historical hit rate. This module's own header scopes historical data
+    // explicitly to the CI ("Wilson CI computed from operator's historical
+    // hit rate... when enough samples exist"); overwriting the point
+    // estimate too collapsed every finding in a CWE×language slice with
+    // enough history to the SAME number regardless of its own reachability/
+    // sanitizer/auth signals — a maximally-dangerous finding and a
+    // maximally-safe one scored identically, while exploitProbabilityWhy
+    // kept listing factors that, in that branch, had zero effect on the
+    // reported number.
     const hist = _historicalCi(history, cwe, f.language || (f.file || '').split('.').pop());
     if (hist) {
-      f.exploitProbability = hist.p;
+      f.exploitProbability = p;
       f.exploitProbabilityCI95 = hist.ci;
       f.exploitProbabilitySlice = hist.slice;
       f.exploitProbabilityN = hist.n;

package/src/posture/falsification.js

@@ -22,7 +22,19 @@ import {
 } from './verification-separation.js';
 
 const DEMOTE_FACTOR = 0.4;               // mirror proof-gate.js
-const TIERS = ['low', 'medium', 'high']; // confidence / exploitability tier order
+// Stage 3 correctness audit (detection depth): this ladder was missing
+// 'critical' — exploitability.js sets f.exploitabilityTier = 'critical' at
+// score >= 0.80 (the tier falsification most needs to demote, since it's
+// exactly the findings a false "survives" verdict would most overstate).
+// _dropTier('critical') hit the `i <= 0` "unknown tier, leave unchanged"
+// branch (indexOf returns -1 for an unrecognized value), so a falsified
+// finding at the critical exploitability tier kept its full tier — the
+// demotion silently no-op'd for the highest tier in the system. 'very-low'
+// (confidence.js's own floor tier) is included too, for the same reason
+// confidenceTier is demoted by this same function — it was already
+// unchanged-at-floor by the same `i <= 0` fallback, so this is a
+// completeness fix there, not a behavior change.
+const TIERS = ['very-low', 'low', 'medium', 'high', 'critical']; // confidence / exploitability tier order
 
 function _dropTier(tier) {
   const i = TIERS.indexOf(tier);
@@ -66,7 +78,16 @@ export function classifyFinding(finding, fileContents) {
   }
   // A sanitizer that doesn't match the sink context does NOT block the flow —
   // the finding survives (this is a real bug, not a mitigation).
-  if (finding.sanitizerMismatch === true) {
+  //
+  // Stage 3 correctness audit (detection depth): this was `=== true`, but
+  // the field's real producer (engine.js's applySanitizerEffectiveness)
+  // sets `f.sanitizerMismatch = f.sanitizerType` — a STRING sanitizer-type
+  // label ("Type Guard", "JWT Algo Pinning", ...), never the literal
+  // boolean `true`. Every OTHER consumer of this field (confidence.js,
+  // exploitability.js, engine.js's own scoring) checks it via plain
+  // truthiness; this strict-equality check could never match a real
+  // finding, making the whole branch dead code.
+  if (finding.sanitizerMismatch) {
     return { verdict: 'survived', reasons: ['wrong-context sanitizer does not neutralize this sink'] };
   }
   const window = _pathWindow(finding, fileContents);

package/src/posture/fix-metrics.js

@@ -62,7 +62,7 @@ export function recordFixAttempt(scanRoot, record) {
   try {
     const dir = stateDir(scanRoot);
     if (!isSafeStateDir(dir)) return false;
-    if (!stateWritesEnabled()) return;
+    if (!stateWritesEnabled()) return false;
   fs.mkdirSync(dir, { recursive: true });
     // One writeSync of one newline-terminated line: a concurrent reader sees
     // whole records or nothing, and a torn tail is dropped on read.

package/src/posture/fix-verify-loop.js

@@ -38,7 +38,16 @@ function _detectRunner(scanRoot) {
     } catch { return null; }
   })();
   if (pkg && pkg.scripts && pkg.scripts.test && !/no test specified/.test(String(pkg.scripts.test))) {
-    return { runner: 'npm', cmd: 'npm', args: ['test', '--silent', '--', '--passWithNoTests'] };
+    // --passWithNoTests is Jest-specific CLI syntax — appending it
+    // unconditionally broke every non-Jest npm test script (mocha, vitest,
+    // ava, tap, or a plain node script) with an "unrecognized option" exit,
+    // failing verification for a reason that has nothing to do with
+    // whether the patch actually broke anything. Only add it when Jest is
+    // actually the configured runner.
+    const usesJest = /\bjest\b/.test(String(pkg.scripts.test))
+      || Boolean(pkg.devDependencies?.jest) || Boolean(pkg.dependencies?.jest);
+    const args = usesJest ? ['test', '--silent', '--', '--passWithNoTests'] : ['test', '--silent'];
+    return { runner: 'npm', cmd: 'npm', args };
   }
   if (has('pytest.ini') || has('pyproject.toml') || has('setup.cfg')) {
     return { runner: 'pytest', cmd: 'pytest', args: ['-q', '--no-header', '-x'] };