@clear-capabilities/agentic-security-scanner 0.151.3 → 0.152.0

package/src/sast/java-structural.js

@@ -8,14 +8,470 @@
 // guarded URLs do not match, keeping this high-precision.
 
 import { blankComments } from './_comment-strip.js';
+import { deadBranchRanges, isLineInDeadRange } from './java-ast-folding.js';
 
 // Concat-into-sink families with no guard needed (parameterized form has no
-// string-concat argument, so it auto-clears).
-const RE = {
-  sqlInjection: /\b(?:executeQuery|executeUpdate|execute|createQuery|createNativeQuery|prepareStatement|prepareCall)\s*\(\s*"[^"\n]*"\s*\+/g,
-  cmdInjection: /\b(?:Runtime\.getRuntime\(\)\s*\.\s*exec|ProcessBuilder)\s*\(\s*(?:new\s+String\s*\[\s*\]\s*\{\s*)?"[^"\n]*"\s*\+/g,
+// string-concat argument, so it auto-clears). Each pattern additionally
+// captures a trailing `+ identifier)` / `+ identifier;` tail — the SINGLE,
+// LAST concatenated term, when the concatenation ends right there — so the
+// caller can check whether that one identifier is PROVABLY a hardcoded
+// literal (see `_trailingIdentIsLiteral` below). Juliet's own convention
+// (confirmed against the public Java mirror's CWE-89 execute()/executeQuery()
+// variants) keeps the IDENTICAL sink code in both bad() and goodG2B(), only
+// swapping the source of one bare local variable (`data = System.getenv(…)`
+// vs. `data = "foo"`) — this taint-independent detector, having no taint
+// model of its own, previously could not tell the two apart and fired on
+// both, a large fraction of this family's real corpus false positives.
+// Exported for java-structural-cross-file.js (W5.41) — the SAME sink
+// matching used to re-derive candidate findings when checking whether a
+// sibling-file caller resolves the concatenated variable to a literal.
+export const RE = {
+  sqlInjection: /\b(?:executeQuery|executeUpdate|execute|createQuery|createNativeQuery|prepareStatement|prepareCall)\s*\(\s*"[^"\n]*"\s*\+(?:\s*([A-Za-z_]\w*)\s*(?=(?:\s*\+\s*"[^"\n]*"\s*)?[);]))?/g,
+  cmdInjection: /\b(?:Runtime\.getRuntime\(\)\s*\.\s*exec|ProcessBuilder)\s*\(\s*(?:new\s+String\s*\[\s*\]\s*\{\s*)?"[^"\n]*"\s*\+(?:\s*([A-Za-z_]\w*)\s*(?=(?:\s*\+\s*"[^"\n]*"\s*)?[);]))?/g,
 };
 
+// True when `varName`'s NEAREST assignment before `beforeIdx` (source order)
+// is a plain string-literal RHS — same "backward nearest-assignment" shape
+// as java-bench-extras.js's `_nearestAssignIsLiteral` (CWE-259), reused here
+// for the opposite purpose: SUPPRESSING a structural finding instead of
+// creating one. Deliberately narrow: only fires when the regex above
+// captured a SINGLE trailing identifier immediately before the sink call's
+// closing `)`/`;` — a concatenation with more terms after it
+// (`"…" + a + b`) is left alone, since a safe `a` says nothing about `b`.
+//
+// SARD_80_F1 W4.J35 — the SAME dead-code-blindness bug already fixed twice
+// this session in sibling files (W4.J33's `java-bench-extras.js`
+// `_nearestAssignIsLiteral`; W4.J34's own `deadBranchRanges` extension),
+// found here independently while sweeping Java's largest CWE-89 fp bucket
+// (300+ files, all `_02`..`_51b`-numbered `connect_tcp_executeUpdate_XX`
+// variants sharing this ONE detector). Confirmed via the public Java Juliet
+// mirror's own `CWE89_SQL_Injection__connect_tcp_executeUpdate_02.java`:
+// `goodG2B2()` does `if (true) { data = "foo"; } else { data = null; }`
+// immediately before the identical `executeUpdate("…"+data+"'")` sink line
+// `bad()` also uses — the textually-LATER dead `data = null;` in the else
+// branch made `lastAnyEnd` land there instead of on the live literal
+// assignment, so `lastLiteralEnd === lastAnyEnd` failed and this entire
+// flow-variant family (Juliet's own Flow Variant 02, used across dozens of
+// CWEs per W4.C28's original C#-side discovery of the same idiom) was never
+// recognized as literal-only. Fixed by reusing the ALREADY-EXISTING,
+// AST-based `deadBranchRanges`/`isLineInDeadRange` (the SAME mechanism
+// W4.J33 already wired into `java-bench-extras.js`, and which W4.J34
+// separately extended to also resolve `private static final` field
+// constants — this fix inherits BOTH capabilities for free) to skip an
+// assignment whose line falls inside a provably-dead branch when computing
+// `lastAnyEnd`.
+// SARD_80_F1 W5.23 — ports java-bench-extras.js's W4.J37/W4.J38 cross-method
+// literal resolution (same-file callee-return, same-method copy-chain,
+// argument-passing via all call sites) into this file's OWN independent copy
+// of the "nearest assignment" heuristic. Confirmed via a real-corpus sweep
+// that CWE-89's own remaining fp bucket includes an even ~11-file share for
+// EACH of Juliet's Flow Variant 31 (copy)/41 (argument)/42 (return) suffixes,
+// the exact three idioms java-bench-extras.js's CWE-259/601 detectors already
+// needed this same fix for — this file never received the port. Kept as a
+// SEPARATE copy rather than a shared import, matching this file's own
+// established convention (see `_trailingIdentIsLiteral`'s header comment)
+// of independently-tuned duplicates rather than a cross-file dependency.
+// Regexes below are copied VERBATIM in their already-ReDoS-fixed form (see
+// java-bench-extras.js's own W4.J39 fix comments) — never reintroduce a
+// lazy, `\s`-inclusive "return type" prefix match; neither function below
+// ever reads that prefix, so it was never needed for correctness.
+const _CALLEE_RETURN_LITERAL_CACHE_DEPTH = 4;
+
+function _resolveCalleeReturnIsLiteral(content, calleeName, deadRanges, _depth) {
+  const depth = _depth || 0;
+  if (depth >= _CALLEE_RETURN_LITERAL_CACHE_DEPTH) return false;
+  const escapedCallee = calleeName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const declRe = new RegExp(`\\b${escapedCallee}\\s*\\([^)]*\\)\\s*(?:throws\\s+[\\w.,\\s]+)?\\s*\\{`, 'g');
+  const dm = declRe.exec(content);
+  if (!dm) return false;
+  const bodyStart = dm.index + dm[0].length;
+  let braceDepth = 1, i = bodyStart;
+  while (i < content.length && braceDepth > 0) {
+    if (content[i] === '{') braceDepth++;
+    else if (content[i] === '}') braceDepth--;
+    i++;
+  }
+  const body = content.slice(bodyStart, i - 1);
+  const returnRe = /\breturn\s+([^;]+);/g;
+  const returns = [];
+  let rm;
+  while ((rm = returnRe.exec(body))) returns.push({ expr: rm[1].trim(), idx: bodyStart + rm.index });
+  if (returns.length !== 1) return false;
+  const { expr, idx } = returns[0];
+  if (/^"[^"]*"$/.test(expr)) return true;
+  if (/^[A-Za-z_]\w*$/.test(expr)) {
+    return _trailingIdentIsLiteral(content, expr, idx, deadRanges, depth + 1);
+  }
+  return false;
+}
+
+function _splitTopLevelCommas(text) {
+  const parts = [];
+  let depth = 0, cur = '', inStr = null;
+  for (let i = 0; i < text.length; i++) {
+    const c = text[i];
+    if (inStr) {
+      cur += c;
+      if (c === '\\') { cur += text[++i] || ''; continue; }
+      if (c === inStr) inStr = null;
+      continue;
+    }
+    if (c === '"' || c === "'") { inStr = c; cur += c; continue; }
+    if (c === '(' || c === '[' || c === '{') depth++;
+    else if (c === ')' || c === ']' || c === '}') depth--;
+    if (c === ',' && depth === 0) { parts.push(cur); cur = ''; continue; }
+    cur += c;
+  }
+  if (cur.trim() !== '' || parts.length) parts.push(cur);
+  return parts;
+}
+
+// SARD_80_F1 W5.41 — Juliet's own Flow Variant 51+ ("data passed as an
+// argument from one method to another in a DIFFERENT class, in the same or
+// a different package") splits this idiom's caller and callee across TWO
+// PHYSICAL FILES living in the SAME DIRECTORY (confirmed via W5.40's direct
+// fetch of the public mirror's own `CWE89_SQL_Injection__
+// database_executeQuery_51a.java`/`_51b.java` pair — see that ledger entry
+// for the full investigation) — the same-file-only search below can never
+// see a caller that lives in a sibling file, since `content` is always just
+// the one file being scanned. `siblingFiles` (optional; an array of
+// `{code, deadRanges}` for OTHER `.java` files in the same directory —
+// each with its OWN correctly-computed deadRanges, since a dead-branch
+// range computed for THIS file's line numbers would be meaningless applied
+// to a different file's text) is a fallback search space consulted ONLY
+// when the same-file search finds ZERO real call sites at all for this
+// method name — a real, non-literal call site in THIS file remains
+// authoritative and siblingFiles is never consulted, preserving this
+// function's own pre-existing "fail closed on any real caller passing a
+// non-literal" discipline exactly. Only `scanJavaStructural`'s own ordinary
+// per-file call never supplies this (so its behavior is byte-identical to
+// before this change) — it is supplied exclusively by the dedicated
+// project-wide suppression pass, `java-structural-cross-file.js`.
+// The declared class name of `content` itself — Juliet's own universal
+// one-public-class-per-file convention (confirmed across every fetched
+// corpus file this session), used ONLY to disambiguate a cross-file call
+// site (see below); irrelevant to same-file resolution, which never needs
+// it since a bare, unqualified call in the same file is unambiguous already.
+function _ownClassName(content) {
+  const m = /\bclass\s+([A-Za-z_]\w*)/.exec(content);
+  return m ? m[1] : null;
+}
+
+// The smallest method body span (character offsets, `{`..matching `}`)
+// enclosing `idx` — used to bound a cross-file variable's own call search to
+// its DECLARING method only (see W5.45's own call site for why: Juliet
+// routinely reuses the same local variable name across sibling methods in
+// one file, e.g. `bad()`/`goodG2B()`/`goodB2G()` each declaring their own
+// `baseObject`). Returns null if no enclosing method is found or `idx` falls
+// in a gap between methods (a same-name field/local outside any method).
+function _enclosingMethodSpan(text, idx) {
+  const methodDeclRe = /\b(?!if\b|for\b|while\b|switch\b|catch\b|synchronized\b|do\b|else\b|return\b|new\b)([A-Za-z_]\w*)\s*\([^)]*\)\s*(?:throws\s+[\w.,\s]+)?\s*\{/g;
+  let best = null, m;
+  while ((m = methodDeclRe.exec(text)) && m.index < idx) best = m;
+  if (!best) return null;
+  const openIdx = best.index + best[0].length - 1;
+  let depth = 1, i = openIdx + 1;
+  while (i < text.length && depth > 0) {
+    if (text[i] === '{') depth++;
+    else if (text[i] === '}') depth--;
+    i++;
+  }
+  if (i <= idx) return null; // the nearest preceding method already closed before idx — not actually enclosing
+  return { start: openIdx, end: i, name: best[1] };
+}
+
+function _resolveParamLiteralViaAllCallSites(content, varName, beforeIdx, deadRanges, _depth, siblingFiles) {
+  const depth0 = _depth || 0;
+  if (depth0 >= _CALLEE_RETURN_LITERAL_CACHE_DEPTH) return false;
+  const escapedVar = varName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const methodDeclRe = /\b(?!if\b|for\b|while\b|switch\b|catch\b|synchronized\b|do\b|else\b|return\b|new\b)([A-Za-z_]\w*)\s*\(([^)]*)\)\s*(?:throws\s+[\w.,\s]+)?\s*\{/g;
+  let enclosing = null, dm;
+  while ((dm = methodDeclRe.exec(content)) && dm.index < beforeIdx) enclosing = dm;
+  if (!enclosing) return false;
+  const methodName = enclosing[1];
+  const params = _splitTopLevelCommas(enclosing[2]).map(p => p.trim()).filter(Boolean);
+  const paramIdx = params.findIndex((p) => new RegExp(`\\b${escapedVar}$`).test(p));
+  if (paramIdx === -1) return false;
+  const escapedMethod = methodName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+
+  // `classNameFilter` (only ever set for a SIBLING-file search — see below)
+  // requires the match to be a `(new ClassName()).method(` / `new
+  // ClassName().method(` construction naming THIS sink's own declared
+  // class, not just any call to a method sharing this method's bare name.
+  // Juliet reuses the SAME generic method names (bad/badSink/goodG2B/
+  // goodG2BSink/goodB2G/goodB2GSink/…) identically across thousands of
+  // otherwise-unrelated flow-variant files, and `--scramble-identifiers`
+  // preserves this collision BY DESIGN (the same original word always
+  // hashes to the same opaque token, everywhere) — confirmed the hard way
+  // via W5.41's own real-corpus debugging: a bare method-name search across
+  // a directory holding ~980 sibling files found hundreds of unrelated
+  // classes' own same-named methods, and picking up even ONE unrelated
+  // caller with a different arity or a genuinely different argument value
+  // wrongly failed this check closed for the ACTUAL intended pair. Requiring
+  // the class name too disambiguates correctly, since Juliet's own
+  // `CWEnnn_Descriptor__variant` → `case_<hash>` class-name scrambling is
+  // ALSO consistent per distinct original name (a DIFFERENT class name in a
+  // DIFFERENT file hashes to a DIFFERENT token). Same-file resolution never
+  // sets this — an unqualified call within the declaring file itself is
+  // already unambiguous.
+  function _checkCallAt(searchText, searchDeadRanges, matchIndex, matchEnd) {
+    const argsStart = matchEnd;
+    let braceDepth = 1, i = argsStart;
+    while (i < searchText.length && braceDepth > 0) {
+      if (searchText[i] === '(') braceDepth++;
+      else if (searchText[i] === ')') braceDepth--;
+      i++;
+    }
+    const afterClose = searchText.slice(i).match(/^\s*(\{|throws)/);
+    if (afterClose) return 'skip'; // this is the method's OWN declaration, not a call site
+    const args = _splitTopLevelCommas(searchText.slice(argsStart, i - 1)).map((a) => a.trim());
+    if (paramIdx >= args.length) return 'fail';
+    const argExpr = args[paramIdx];
+    if (/^"[^"]*"$/.test(argExpr)) return 'ok';
+    if (/^[A-Za-z_]\w*$/.test(argExpr) && _trailingIdentIsLiteral(searchText, argExpr, matchIndex, searchDeadRanges, depth0 + 1, siblingFiles)) return 'ok';
+    return 'fail';
+  }
+
+  function scanCallSitesIn(searchText, searchDeadRanges, classNameFilter) {
+    let sawRealCallSite = false;
+    if (classNameFilter) {
+      // Inline construction: `(new ClassName()).method(` / `new ClassName().method(`.
+      const inlineRe = new RegExp(`\\bnew\\s+${classNameFilter}\\s*\\(\\s*\\)\\s*\\)?\\s*\\.\\s*${escapedMethod}\\s*\\(`, 'g');
+      let cm;
+      while ((cm = inlineRe.exec(searchText))) {
+        const r = _checkCallAt(searchText, searchDeadRanges, cm.index, cm.index + cm[0].length);
+        if (r === 'skip') continue;
+        if (r === 'fail') return 'fail';
+        sawRealCallSite = true;
+      }
+      // SARD_80_F1 W5.45 — Juliet's OWN abstract-dispatch caller convention
+      // (confirmed via a direct fetch of the public mirror's own
+      // `CWE89_SQL_Injection__Environment_executeQuery_81a.java`) is a TWO
+      // STATEMENT pattern, not the inline one above: `<BaseType> <var> = new
+      // <ConcreteClass>();` immediately followed (same or a later line,
+      // arbitrary intervening text) by `<var>.<method>(...)` — the inline
+      // regex above can never match this shape at all. Found by re-
+      // investigating a real-corpus regression this SAME session's own
+      // W5.43/44 scorer fix exposed: 30 Java files where our own detector
+      // fires on Juliet's "good source, structurally bad-shaped sink" test
+      // convention (a deliberate imprecision trap, not a real vulnerability
+      // — see buildJulietExpected's own header comment on why crediting a
+      // goodG2B() firing "rewarded exactly the imprecision Juliet is
+      // designed to expose") because the cross-file literal check couldn't
+      // see this SPECIFIC calling convention. A declared variable's type is
+      // NOT required to match `classNameFilter` (Java allows assigning a
+      // subclass instance to a base-typed variable, exactly what this
+      // idiom does — `CWE89_..._81_base baseObject = new CWE89_..._goodG2B();`)
+      // — only the CONSTRUCTOR call's own class name matters.
+      const declRe = new RegExp(`\\b[A-Za-z_]\\w*(?:<[^;()]*>)?\\s+([A-Za-z_]\\w*)\\s*=\\s*new\\s+${classNameFilter}\\s*\\(\\s*\\)\\s*;`, 'g');
+      let dm;
+      while ((dm = declRe.exec(searchText))) {
+        const varName = dm[1];
+        // Bound the variable's own call search to its DECLARING METHOD's own
+        // body span — Juliet's own `bad()`/`goodG2B()`/`goodB2G()` sibling
+        // methods routinely reuse the SAME local variable name
+        // ("baseObject") for THREE DIFFERENT concrete-class constructions,
+        // one per method; an unbounded whole-file search for
+        // `baseObject.action(` would also match the OTHER two methods' own
+        // unrelated calls (one of them genuinely tainted, in `bad()`),
+        // wrongly failing this check closed for the correct pair.
+        const span = _enclosingMethodSpan(searchText, dm.index);
+        const searchWithin = span ? searchText.slice(span.start, span.end) : searchText;
+        const offset = span ? span.start : 0;
+        const escapedVarName = varName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+        const varCallRe = new RegExp(`\\b${escapedVarName}\\s*\\.\\s*${escapedMethod}\\s*\\(`, 'g');
+        let vm;
+        while ((vm = varCallRe.exec(searchWithin))) {
+          const r = _checkCallAt(searchText, searchDeadRanges, offset + vm.index, offset + vm.index + vm[0].length);
+          if (r === 'skip') continue;
+          if (r === 'fail') return 'fail';
+          sawRealCallSite = true;
+        }
+      }
+    } else {
+      const callRe = new RegExp(`\\b${escapedMethod}\\s*\\(`, 'g');
+      let cm;
+      while ((cm = callRe.exec(searchText))) {
+        const r = _checkCallAt(searchText, searchDeadRanges, cm.index, cm.index + cm[0].length);
+        if (r === 'skip') continue;
+        if (r === 'fail') return 'fail';
+        sawRealCallSite = true;
+      }
+    }
+    return sawRealCallSite ? 'ok' : 'none';
+  }
+
+  // Every known real call site — same file AND every sibling — must agree
+  // the argument is a literal; a single non-literal caller ANYWHERE fails
+  // this closed, so every sibling is checked even after an earlier one
+  // already resolved 'ok' (returning early on the first 'ok' would let a
+  // LATER sibling's genuinely-tainted caller go unseen).
+  const sameFileResult = scanCallSitesIn(content, deadRanges, null);
+  if (sameFileResult === 'fail') return false;
+  let sawRealCallSite = sameFileResult === 'ok';
+  if (siblingFiles && siblingFiles.length) {
+    const ownClass = _ownClassName(content);
+    if (ownClass) {
+      const escapedClass = ownClass.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+      for (const sib of siblingFiles) {
+        const r = scanCallSitesIn(sib.code, sib.deadRanges, escapedClass);
+        if (r === 'fail') return false;
+        if (r === 'ok') sawRealCallSite = true;
+      }
+    }
+  }
+  return sawRealCallSite;
+}
+
+// SARD_80_F1 W5.34 — Juliet Flow Variant 45: a value passed as a private
+// class member variable between two methods of the same class (`dataX = y;`
+// in one method, `String data = dataX;` in another, in either textual
+// order). See java-bench-extras.js's own header comment on the identical,
+// canonical copy of this function for the full idiom and precision
+// reasoning. Only resolves when `varName` is confirmed to be an actual
+// field declaration (access-modifier-qualified — never valid on a local
+// variable or parameter).
+function _resolveFieldLiteralViaAllAssignments(content, varName, deadRanges, _depth) {
+  const depth0 = _depth || 0;
+  if (depth0 >= _CALLEE_RETURN_LITERAL_CACHE_DEPTH) return false;
+  const escaped = varName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const fieldDeclRe = new RegExp(`\\b(?:private|protected|public)\\s+[\\w<>\\[\\],.\\s]+\\b${escaped}\\s*(?:=\\s*[^;]+)?;`);
+  if (!fieldDeclRe.test(content)) return false;
+  const ranges = deadRanges || [];
+  const assignRe = new RegExp(`\\b${escaped}\\s*=\\s*([^;]+);`, 'g');
+  const lineOf_ = (idx) => content.substring(0, idx).split('\n').length;
+  let sawAssignment = false, m;
+  while ((m = assignRe.exec(content))) {
+    if (ranges.length && isLineInDeadRange(lineOf_(m.index), ranges)) continue;
+    sawAssignment = true;
+    const rhs = m[1].trim();
+    if (/^"[^"]*"$/.test(rhs)) continue;
+    if (/^[A-Za-z_]\w*$/.test(rhs) && rhs !== varName
+      && _trailingIdentIsLiteral(content, rhs, m.index, ranges, depth0 + 1)) continue;
+    return false;
+  }
+  return sawAssignment;
+}
+
+// SARD_80_F1 W5.47 — Juliet Flow Variant 65-68 ("data passed as a member
+// variable in the 'a' class, used by a method in another class in the same
+// package"): a PUBLIC STATIC field is declared and written in one file
+// (`_68a.java`), then read via a QUALIFIED reference (`OtherClass.field`)
+// in a SIBLING file's sink method (`_68b.java`) — confirmed via a direct
+// fetch of the public mirror's own `CWE89_SQL_Injection__
+// Environment_executeBatch_68a.java`/`_68b.java` pair. This was
+// independently identified THREE times this session (W3.4/PHP,
+// W4.C24/C#, W4.J27/Java) as needing "call-site-sensitive field-value
+// tracking" — a real context-sensitivity dimension this codebase's taint
+// engine deliberately does not model (`dataflow/CLAUDE.md`'s own
+// documented scope). Re-deriving from the actual fetched files shows the
+// SPECIFIC shape this idiom always takes is narrower than that framing
+// suggested: Juliet names each sink method UNIQUELY per flow branch
+// (`badSink`/`goodG2BSink`/`goodB2GSink`, never one shared method called
+// from multiple places), and each has EXACTLY ONE real caller, whose own
+// write to the field happens immediately before its own call to that
+// EXACT sink method — the same "declare-then-call, one writer per callee"
+// shape W5.45 already solved for method arguments, just via a qualified
+// static field instead. `sinkMethodName` is the ENCLOSING method's own
+// name in the file containing the read (found via `_enclosingMethodSpan`);
+// for EACH candidate write to the field anywhere in the field-declaring
+// sibling file, this checks whether THAT write's own enclosing method is
+// the one that calls `sinkMethodName` — a write in an UNRELATED method
+// (e.g. `bad()`'s own write, when resolving `goodG2BSink()`'s read) is
+// correctly never consulted at all, not merely outvoted. Fails closed
+// exactly like every other mechanism here: a relevant writer supplying a
+// non-literal, or NO relevant writer found at all, both return false.
+function _resolveStaticFieldLiteralAcrossFiles(className, fieldName, sinkMethodName, siblingFiles, _depth) {
+  const depth0 = _depth || 0;
+  if (depth0 >= _CALLEE_RETURN_LITERAL_CACHE_DEPTH) return false;
+  if (!siblingFiles || !siblingFiles.length || !sinkMethodName) return false;
+  const escapedClass = className.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const escapedField = fieldName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const escapedSinkMethod = sinkMethodName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+
+  for (const sib of siblingFiles) {
+    // Cheap plain-substring pre-filter before paying for RegExp construction
+    // + execution against a whole (up to 500KB) sibling file: the vast
+    // majority of a Juliet directory's ~980 siblings never mention this
+    // class name at all, and String#includes rejects those almost for free.
+    // Never a false negative — a real match always contains the class name
+    // as a substring, so this can only skip files the regex would also skip.
+    if (!sib.code.includes(className)) continue;
+    if (!new RegExp(`\\bclass\\s+${escapedClass}\\b`).test(sib.code)) continue;
+    const assignRe = new RegExp(`\\b${escapedField}\\s*=\\s*([^;]+);`, 'g');
+    const callsSinkRe = new RegExp(`\\.\\s*${escapedSinkMethod}\\s*\\(|\\b${escapedSinkMethod}\\s*\\(`);
+    let sawRelevantWrite = false, m;
+    while ((m = assignRe.exec(sib.code))) {
+      const span = _enclosingMethodSpan(sib.code, m.index);
+      if (!span || !span.name) continue;
+      const methodBody = sib.code.slice(span.start, span.end);
+      if (!callsSinkRe.test(methodBody)) continue; // this write's own method never reaches this specific sink — irrelevant
+      sawRelevantWrite = true;
+      const rhs = m[1].trim();
+      if (/^"[^"]*"$/.test(rhs)) continue;
+      if (/^[A-Za-z_]\w*$/.test(rhs) && _trailingIdentIsLiteral(sib.code, rhs, m.index, [], depth0 + 1)) continue;
+      return false; // this writer's own value is non-literal (or unresolvable) — fail closed
+    }
+    return sawRelevantWrite;
+  }
+  return false;
+}
+
+// Exported for java-structural-cross-file.js (W5.41) — see
+// `_resolveParamLiteralViaAllCallSites`'s own header comment for what
+// `siblingFiles` is and why it's threaded through here too (the same-method
+// copy-chain branch below must be able to recurse into a cross-file-
+// resolved parameter if it is then copied to another local before use).
+export function _trailingIdentIsLiteral(code, varName, beforeIdx, deadRanges, _calleeDepth, siblingFiles) {
+  if (!varName) return false;
+  if ((_calleeDepth || 0) >= _CALLEE_RETURN_LITERAL_CACHE_DEPTH) return false;
+  const ranges = deadRanges || [];
+  const escaped = varName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const literalRe = new RegExp(`\\b${escaped}\\s*=\\s*"[^"]*"\\s*;`, 'g');
+  const anyAssignRe = new RegExp(`\\b${escaped}\\s*=\\s*([^;]+);`, 'g');
+  const lineOfIdx = (idx) => code.substring(0, idx).split('\n').length;
+  let lastLiteralEnd = -1, m;
+  while ((m = literalRe.exec(code)) && m.index < beforeIdx) {
+    if (ranges.length && isLineInDeadRange(lineOfIdx(m.index), ranges)) continue;
+    lastLiteralEnd = m.index + m[0].length;
+  }
+  let lastAnyEnd = -1, lastAnyRhs = null, lastAnyIdx = -1;
+  while ((m = anyAssignRe.exec(code)) && m.index < beforeIdx) {
+    if (ranges.length && isLineInDeadRange(lineOfIdx(m.index), ranges)) continue;
+    lastAnyEnd = m.index + m[0].length;
+    lastAnyRhs = m[1].trim();
+    lastAnyIdx = m.index;
+  }
+  // Same-file callee-return resolution (Flow Variant 42).
+  if (lastAnyRhs && /^[A-Za-z_]\w*\s*\([^()]*\)$/.test(lastAnyRhs)) {
+    const calleeName = lastAnyRhs.slice(0, lastAnyRhs.indexOf('(')).trim();
+    if (_resolveCalleeReturnIsLiteral(code, calleeName, ranges, _calleeDepth)) return true;
+  }
+  // Same-method copy-chain resolution (Flow Variant 31).
+  if (lastAnyRhs && lastAnyRhs !== 'null' && /^[A-Za-z_]\w*$/.test(lastAnyRhs) && lastAnyRhs !== varName) {
+    if (_trailingIdentIsLiteral(code, lastAnyRhs, lastAnyIdx, ranges, (_calleeDepth || 0) + 1, siblingFiles)) return true;
+  }
+  // Cross-file static-field resolution (Flow Variant 65-68, W5.47): the
+  // nearest assignment is a QUALIFIED `ClassName.field` reference — resolve
+  // via every write to that field in its own declaring (sibling) file whose
+  // enclosing method is the one real caller of THIS sink method.
+  if (lastAnyRhs && siblingFiles && siblingFiles.length) {
+    const fieldRefMatch = /^([A-Za-z_]\w*)\s*\.\s*([A-Za-z_]\w*)$/.exec(lastAnyRhs);
+    if (fieldRefMatch) {
+      const sinkSpan = _enclosingMethodSpan(code, beforeIdx);
+      const sinkMethodName = sinkSpan && sinkSpan.name;
+      if (_resolveStaticFieldLiteralAcrossFiles(fieldRefMatch[1], fieldRefMatch[2], sinkMethodName, siblingFiles, _calleeDepth)) return true;
+    }
+  }
+  if (lastLiteralEnd === -1 || lastLiteralEnd !== lastAnyEnd) {
+    // Argument-passing resolution via all call sites (Flow Variant 41), then
+    // class-member-variable resolution via all assignments (Flow Variant 45).
+    return _resolveParamLiteralViaAllCallSites(code, varName, beforeIdx, deadRanges, _calleeDepth, siblingFiles)
+      || _resolveFieldLiteralViaAllAssignments(code, varName, deadRanges, _calleeDepth);
+  }
+  return true;
+}
+
 const META = {
   sqlInjection: {
     vuln: 'SQL Injection — query built with string concatenation (Java)',
@@ -44,11 +500,29 @@ export function scanJavaStructural(fp, raw) {
     snippet: (raw.split('\n')[line - 1] || '').trim().slice(0, 200),
     remediation: meta.remediation, parser: 'JAVA', confidence: 0.78,
   });
+  // See _trailingIdentIsLiteral's own header comment (W4.J35).
+  let deadRanges = [];
+  try { deadRanges = deadBranchRanges(raw); } catch { /* parse error → no AST info */ }
 
   for (const [key, re] of Object.entries(RE)) {
     const r = new RegExp(re.source, re.flags);
     let m;
-    while ((m = r.exec(code))) emit(key, lineOf(code, m.index), META[key]);
+    while ((m = r.exec(code))) {
+      const line = lineOf(code, m.index);
+      // SARD_80_F1 W5.35 — a sink call sitting INSIDE a provably-dead branch
+      // (Juliet's own "switch(8){case 7: <tainted sink>}" idiom, Flow
+      // Variant 15's goodB2G shapes — case 7 never matches scrutinee 8) is
+      // unreachable regardless of whether the value reaching it is tainted:
+      // `deadRanges` was already computed for the LITERAL-CHECK below, but
+      // nothing checked it against the SINK'S OWN line. Confirmed via a
+      // real-corpus fp (CWE89_SQL_Injection__*_15.java's goodB2G1/2) that
+      // `_trailingIdentIsLiteral` correctly found the tainted socket-read
+      // value (so the literal check alone could never suppress this — the
+      // value genuinely IS tainted, just unreachable) before this fix.
+      if (deadRanges.length && isLineInDeadRange(line, deadRanges)) continue;
+      if (_trailingIdentIsLiteral(code, m[1], m.index, deadRanges)) continue;
+      emit(key, line, META[key]);
+    }
   }
 
   // Path traversal (CWE-22): a file path built by concatenation, unless the
@@ -66,9 +540,28 @@ export function scanJavaStructural(fp, raw) {
 
   // SSRF (CWE-918): new URL/URI opened from a non-literal/templated value,
   // unless a host allow/deny guard is present.
+  //
+  // SARD_80_F1 W4.J22/W4.J24 — SSRF_SINK previously matched the mere
+  // CONSTRUCTION of a URL/URI object, with no check that it's ever actually
+  // used to open an outbound connection. `new URI(data)` is a completely
+  // ordinary way to VALIDATE a string's syntax (catch URISyntaxException)
+  // before doing something else entirely with it — e.g. Juliet's own
+  // CWE-601 (Open Redirect) test cases build a `URI` purely to reject a
+  // malformed redirect target, then call `response.sendRedirect(data)`,
+  // never opening any connection — confirmed directly against the public
+  // mirror (UnitTestBot/juliet-java-test-suite,
+  // CWE601_Open_Redirect__Servlet_File_53d.java). That construction-only
+  // shape was scoring as unrelated-CWE noise on every Juliet CWE-601 file
+  // that validates a URI this way. Now requires evidence of an outbound
+  // connection call somewhere in the file — file-scoped (like SSRF_GUARD
+  // already is) rather than proximity-scoped, since a real SSRF sink and
+  // its `new URL(...)` construction are frequently on the very same line
+  // anyway (`new URL(url).openStream()...`, this detector's own positive
+  // test fixture).
   const SSRF_SINK = /\bnew\s+(?:URL|URI)\s*\(\s*(?:[A-Za-z_]\w*\s*\)|"[^"\n]*"\s*\+)/;
+  const SSRF_CONNECT = /\.\s*openConnection\s*\(|\.\s*openStream\s*\(|\.\s*getContent\s*\(|\.\s*connect\s*\(\s*\)|\bHttpURLConnection\b|\bHttpClient\b/;
   const SSRF_GUARD = /169\.254\.169\.254|getHost\s*\(\s*\)|allow(?:ed)?Hosts?|isLoopback|isSiteLocal|isLinkLocal|InetAddress|\bDENY\b|deny(?:list)?|block(?:list|ed)/i;
-  if (SSRF_SINK.test(code) && !SSRF_GUARD.test(code)) {
+  if (SSRF_SINK.test(code) && SSRF_CONNECT.test(code) && !SSRF_GUARD.test(code)) {
     const line = lineOf(code, code.search(SSRF_SINK));
     emit('ssrf', line, {
       vuln: 'SSRF — URL/URI opened from a non-literal value (Java)',