slopbrick 0.21.0 → 0.21.2

package/README.md

@@ -46,7 +46,7 @@ on every scan — your repository, encoded for the next agent.
   lib, modal system, API client) once. The agent and the linter
   enforce it together.
 
-**Status:** v0.20.0 (current). See the [CHANGELOG](./CHANGELOG.md) for
+**Status:** v0.21.0 (current). See the [CHANGELOG](./CHANGELOG.md) for
 the full release notes.
 
 ---
@@ -76,21 +76,25 @@ For every other config question, see [`EXAMPLES.md`](./EXAMPLES.md).
 
 ---
 
-## The headlines (4-score model, v0.20.0+)
+## The headlines (4-score model, v0.21.0+)
 
-> **v0.15.0 introduced the 4-score model; v0.16.0 + v0.17.0 completed it.**
-> The single `Slop Index` is replaced by **4 independent scores**
-> (all 0-100, **higher = better**). The legacy `slopIndex` field
-> is kept as optional on `ProjectReport` for backward compat with
+> **v0.15.0 introduced the 4-score model; v0.21.0 FLIPPED `aiSlopScore`
+> to the natural-reading "raw amount" direction (0=clean, 100=saturated).**
+> The other three scores stay "higher = better". The legacy `slopIndex`
+> field is kept as optional on `ProjectReport` for backward compat with
 > existing test fixtures and historical telemetry; the v0.14-compat
 > removal is tracked separately.
 
-| Score | What it measures | CI gate? |
-|-------|------------------|----------|
-| **`aiSlopScore`** | AI-slop signatures (16 `ai/*` rules). Raw amount of slop (0=clean, 100=saturated). | **Yes** (≤ `meanSlop: 30` passes) |
-| **`engineeringHygiene`** | Average of 6 category scores: arch, logic, layout, visual, component, test | No (informational) |
-| **`security`** | AI Security Risk band: low=100, medium=67, high=33, critical=0 | No (informational) |
-| **`repositoryHealth`** (composite) | Weighted: 0.4×aiQ + 0.3×eng + 0.2×sec + 0.1×test | No (informational) |
+| Score | What it measures | Direction | CI gate? |
+|-------|------------------|-----------|----------|
+| **`aiSlopScore`** | AI-slop signatures (16 `ai/*` rules). | **lower = cleaner** (raw amount) | **Yes** (`≤ meanSlop: 30` passes) |
+| **`engineeringHygiene`** | Average of 6 category scores: arch, logic, layout, visual, component, test | higher = better | No (informational) |
+| **`security`** | AI Security Risk band: low=100, medium=67, high=33, critical=0 | higher = better | No (informational) |
+| **`repositoryHealth`** (composite) | Weighted: `0.4 × (100 − aiSlopScore) + 0.3 × eng + 0.2 × sec + 0.1 × test` | higher = better (inverts `aiSlopScore` internally) | No (informational) |
+
+**Score-band messages** (v0.21.0+): every score ships with a one-line
+verdict in the pretty output — e.g. `AI Slop Score: 25 → "low amount
+of slop"`, `Security Risk: 33 → "high risk"`. See `src/report/pretty.ts`.
 
 The same numbers are in `.slopbrick/health.json`.
 

package/dist/engine/worker.cjs

@@ -6424,6 +6424,7 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
   if (source) {
     const { line, column } = positionFrom(node, vctx.lineOffsets);
     const importedNames = [];
+    const isTypeOnly = node.typeOnly === true;
     const specifiers = node.specifiers;
     if (Array.isArray(specifiers)) {
       for (const specifier of specifiers) {
@@ -6439,7 +6440,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           }
         } else if (specifier.type === "ImportSpecifier") {
@@ -6454,7 +6456,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           } else if (isObject(local) && local.type === "Identifier" && typeof local.value === "string") {
             const name = local.value;
@@ -6465,7 +6468,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           }
         }
@@ -6758,7 +6762,7 @@ function handleJSXOpeningElement(node, _parent, _path, vctx) {
   }
   return false;
 }
-function handleVariableDeclarator(node, _parent, path, vctx) {
+function handleVariableDeclarator(node, parent, path, vctx) {
   if (!isObject(node)) return false;
   const init = node.init;
   if (init && vctx.visit) {
@@ -6773,16 +6777,18 @@ function handleVariableDeclarator(node, _parent, path, vctx) {
     }
   }
   if (bindingNames.length > 0) {
-    const parent = node.parent;
-    const kind = isObject(parent) && parent.type === "VariableDeclaration" ? String(parent.kind ?? "var") : "var";
+    const declParent = isObject(parent) && parent.type === "ExportNamedDeclaration" ? parent.declaration ?? parent : parent;
+    const kind = isObject(declParent) && declParent.type === "VariableDeclaration" ? String(declParent.kind ?? "var") : "var";
     const { line, column } = positionFrom(id, vctx.lineOffsets);
+    const scope = vctx.ctx.stack.length === 0 ? "module" : "function";
     for (const bindingName of bindingNames) {
       vctx.facts.deadCode.bindings.push({
         name: bindingName,
         kind,
         line,
         column,
-        isReferenced: false
+        isReferenced: false,
+        scope
       });
     }
   }
@@ -33850,6 +33856,16 @@ var aiSegmentSurprisalCvRule = createRule({
   category: "ai",
   severity: "medium",
   aiSpecific: true,
+  // v0.21.0: defaultOff. The rule fires on 250/250 files in the
+  // self-scan (1 per file), which means it's not discriminating
+  // between AI and non-AI text — it fires on every file. The
+  // Binoculars-style per-segment entropy CV is too coarse a
+  // signal at the file level; the entropy floor is hit by most
+  // structured TypeScript regardless of whether the author was
+  // an LLM. Re-enable per-file via
+  // `rules: { 'ai/segment-surprisal-cv': 'medium' }` in
+  // slopbrick.config.mjs once a tighter threshold is calibrated.
+  defaultOff: true,
   description: "Per-segment cross-entropy CV is suspiciously low \u2014 Binoculars (Hans 2024): AI text has near-constant per-window entropy",
   create(context) {
     return context;
@@ -34990,6 +35006,7 @@ var unusedImportRule = createRule({
       if (binding.kind !== "import-specifier" && binding.kind !== "import-default" && binding.kind !== "import-namespace") {
         continue;
       }
+      if (binding.isTypeOnly) continue;
       if (binding.isReferenced) continue;
       if (!binding.name) continue;
       const source = binding.source ? ` from '${binding.source}'` : "";
@@ -35063,6 +35080,9 @@ var unusedLocalRule = createRule({
       if (binding.isReferenced) continue;
       if (SKIP_NAMES.has(binding.name)) continue;
       if (binding.name.startsWith("_")) continue;
+      if (binding.scope === "module" && (binding.kind === "const" || binding.kind === "function" || binding.kind === "class" || binding.kind === "type" || binding.kind === "interface" || binding.kind === "enum")) {
+        continue;
+      }
       issues.push({
         ruleId: "dead/unused-local",
         category: "logic",
@@ -36059,7 +36079,7 @@ var brokenLinkRule = createRule({
 
 // src/rules/dup/identical-block.ts
 var crypto = __toESM(require("crypto"), 1);
-var WINDOW_SIZE = 10;
+var WINDOW_SIZE = 20;
 var MIN_NORMALIZED_LENGTH = 40;
 var HASH_PREFIX_LENGTH = 16;
 var DEDUP_CACHE = /* @__PURE__ */ new Map();
@@ -36284,6 +36304,7 @@ var javaArraylistVsLinkedlistRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     let m;
     NEW_LINKED_LIST_REGEX.lastIndex = 0;
     while ((m = NEW_LINKED_LIST_REGEX.exec(source)) !== null) {
@@ -36318,6 +36339,7 @@ var javaEmptyCatchBlockRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     let m;
     SINGLE_LINE_EMPTY_CATCH_REGEX.lastIndex = 0;
     while ((m = SINGLE_LINE_EMPTY_CATCH_REGEX.exec(source)) !== null) {
@@ -36354,6 +36376,7 @@ var javaLegacyDateApiRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     const flagged = /* @__PURE__ */ new Set();
     let m;
     LEGACY_IMPORT_REGEX.lastIndex = 0;
@@ -36420,6 +36443,7 @@ var javaRawTypeOveruseRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     let m;
     RAW_TYPE_REGEX.lastIndex = 0;
     while ((m = RAW_TYPE_REGEX.exec(source)) !== null) {
@@ -36455,6 +36479,7 @@ var javaStringConcatLoopRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     if (!/\b(for|while|do)\b/.test(source)) return issues;
     let m;
     STRING_CONCAT_REGEX.lastIndex = 0;
@@ -36491,6 +36516,7 @@ var javaSystemOutPrintlnRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     const matches = [];
     let m;
     PRINTLN_REGEX.lastIndex = 0;
@@ -42017,7 +42043,8 @@ var signal_strength_default = {
     precision: 0.6803,
     lastCalibratedAt: "2026-07-01T00:00:00Z",
     verdict: "USEFUL",
-    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=27874, FP=13097, P=68.0%, FPR=5.20%, lift=13.09. v7 was USEFUL (TP=16673, FP=10112, lift=11.29). v8 was USEFUL (TP=11201, FP=2985).",
+    defaultOff: true,
+    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=27874, FP=13097, P=68.0%, FPR=5.20%, lift=13.09. v7 was USEFUL (TP=16673, FP=10112, lift=11.29). v8 was USEFUL (TP=11201, FP=2985). v0.21.0: marked defaultOff because the v0.21.0 self-scan fires 109 times across 47 files \u2014 the heuristic counts `{`, `}`, `(`, `)` as raw characters without skipping string contents, template literals, regex literals, or comments. Files with normal closing-brace tails (the last 4-5 lines of a function definition) consistently fire because the tail has 0 opens and 3-4 closes. The corpus-level 13.09 lift doesn't translate to per-file discrimination. v0.22.0 should re-implement the count with a token-aware state machine (string/comment/regex skip) or switch to a parse-error-based signal.",
     aiSpecific: true,
     _v7Verdict: "USEFUL",
     _v7Lift: 11.29,
@@ -42120,7 +42147,8 @@ var signal_strength_default = {
     precision: 0.7495,
     lastCalibratedAt: "2026-07-01T00:00:00Z",
     verdict: "USEFUL",
-    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=59390, FP=19849, P=75.0%, FPR=7.87%, lift=9.52. v7 was USEFUL (TP=43498, FP=14983, lift=9.11). v8 was USEFUL (TP=15892, FP=4866).",
+    defaultOff: true,
+    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=59390, FP=19849, P=75.0%, FPR=7.87%, lift=9.52. v7 was USEFUL (TP=43498, FP=14983, lift=9.11). v8 was USEFUL (TP=15892, FP=4866). v0.21.0: marked defaultOff because the v0.21.0 self-scan fires 250 times across 250 files (1 per file) \u2014 the Binoculars-style per-segment entropy CV is too coarse a signal at the file level; the entropy floor is hit by most structured TypeScript regardless of whether the author was an LLM. The v8.5 corpus-level lift of 9.52 doesn't translate to per-file discrimination. v0.22.0 should tighten the threshold or re-calibrate against the per-file (not per-corpus) lift.",
     aiSpecific: true,
     _v7Verdict: "USEFUL",
     _v7Lift: 9.11,
@@ -42188,7 +42216,8 @@ var signal_strength_default = {
     precision: 0.5583,
     lastCalibratedAt: "2026-07-01T00:00:00Z",
     verdict: "USEFUL",
-    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=21567, FP=17061, P=55.8%, FPR=6.77%, lift=8.25. v7 was USEFUL (TP=17466, FP=13438, lift=7.71). v8 was USEFUL (TP=4101, FP=3623).",
+    defaultOff: true,
+    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=21567, FP=17061, P=55.8%, FPR=6.77%, lift=8.25. v7 was USEFUL (TP=17466, FP=13438, lift=7.71). v8 was USEFUL (TP=4101, FP=3623). v0.21.2: marked defaultOff because precision 0.5583 < 0.60 floor (the v9.5 methodology in `docs/research/methodology-minimum-sample-size.md` sets the default-on threshold at P >= 0.60). At 55.8% precision, 44% of fires are false positives \u2014 too noisy to gate on. The rule stays available for opt-in via `rules: { 'ai/whitespace-regularity': 'low' }` in slopbrick.config.mjs. The 18 src/ self-scan fires are the typical noise profile.",
     aiSpecific: true,
     _v7Verdict: "USEFUL",
     _v7Lift: 7.71,
@@ -42536,7 +42565,8 @@ var signal_strength_default = {
     precision: 0,
     lastCalibratedAt: "2026-07-01T00:00:00Z",
     verdict: "DORMANT",
-    _calibrationNote: "v0.19: new rule, not yet calibrated. v8.5 calibration does not run dup/* rules. Scheduled for v0.20 calibration on near-dup corpus.",
+    defaultOff: false,
+    _calibrationNote: "v0.19: new rule, not yet calibrated. v8.5 calibration does not run dup/* rules (the calibration script analyzes each file in isolation, but dup/* rules require cross-file state \u2014 the module-scope dedup cache is per-process, not per-file, so per-file analysis never produces cross-file matches and the rule's recall reads as 0). Scheduled for v0.22 calibration on the v9 corpus with a per-project dedup cache. v0.21.1 self-scan: 207 fires in src/ after WINDOW_SIZE 10\u219220 (was 575) \u2014 69% reduction. The actual precision/recall of these 207 fires is unverified ground truth. Default stays on because the rule's value is real (block-level duplications are a classic AI-fingerprint) and the corpus-measured 0.0 is a methodology artifact, not a real measurement.",
     aiSpecific: false,
     _v7Verdict: "DORMANT",
     _v7Lift: 1,

package/dist/engine/worker.js

@@ -6395,6 +6395,7 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
   if (source) {
     const { line, column } = positionFrom(node, vctx.lineOffsets);
     const importedNames = [];
+    const isTypeOnly = node.typeOnly === true;
     const specifiers = node.specifiers;
     if (Array.isArray(specifiers)) {
       for (const specifier of specifiers) {
@@ -6410,7 +6411,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           }
         } else if (specifier.type === "ImportSpecifier") {
@@ -6425,7 +6427,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           } else if (isObject(local) && local.type === "Identifier" && typeof local.value === "string") {
             const name = local.value;
@@ -6436,7 +6439,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           }
         }
@@ -6729,7 +6733,7 @@ function handleJSXOpeningElement(node, _parent, _path, vctx) {
   }
   return false;
 }
-function handleVariableDeclarator(node, _parent, path, vctx) {
+function handleVariableDeclarator(node, parent, path, vctx) {
   if (!isObject(node)) return false;
   const init = node.init;
   if (init && vctx.visit) {
@@ -6744,16 +6748,18 @@ function handleVariableDeclarator(node, _parent, path, vctx) {
     }
   }
   if (bindingNames.length > 0) {
-    const parent = node.parent;
-    const kind = isObject(parent) && parent.type === "VariableDeclaration" ? String(parent.kind ?? "var") : "var";
+    const declParent = isObject(parent) && parent.type === "ExportNamedDeclaration" ? parent.declaration ?? parent : parent;
+    const kind = isObject(declParent) && declParent.type === "VariableDeclaration" ? String(declParent.kind ?? "var") : "var";
     const { line, column } = positionFrom(id, vctx.lineOffsets);
+    const scope = vctx.ctx.stack.length === 0 ? "module" : "function";
     for (const bindingName of bindingNames) {
       vctx.facts.deadCode.bindings.push({
         name: bindingName,
         kind,
         line,
         column,
-        isReferenced: false
+        isReferenced: false,
+        scope
       });
     }
   }
@@ -33821,6 +33827,16 @@ var aiSegmentSurprisalCvRule = createRule({
   category: "ai",
   severity: "medium",
   aiSpecific: true,
+  // v0.21.0: defaultOff. The rule fires on 250/250 files in the
+  // self-scan (1 per file), which means it's not discriminating
+  // between AI and non-AI text — it fires on every file. The
+  // Binoculars-style per-segment entropy CV is too coarse a
+  // signal at the file level; the entropy floor is hit by most
+  // structured TypeScript regardless of whether the author was
+  // an LLM. Re-enable per-file via
+  // `rules: { 'ai/segment-surprisal-cv': 'medium' }` in
+  // slopbrick.config.mjs once a tighter threshold is calibrated.
+  defaultOff: true,
   description: "Per-segment cross-entropy CV is suspiciously low \u2014 Binoculars (Hans 2024): AI text has near-constant per-window entropy",
   create(context) {
     return context;
@@ -34961,6 +34977,7 @@ var unusedImportRule = createRule({
       if (binding.kind !== "import-specifier" && binding.kind !== "import-default" && binding.kind !== "import-namespace") {
         continue;
       }
+      if (binding.isTypeOnly) continue;
       if (binding.isReferenced) continue;
       if (!binding.name) continue;
       const source = binding.source ? ` from '${binding.source}'` : "";
@@ -35034,6 +35051,9 @@ var unusedLocalRule = createRule({
       if (binding.isReferenced) continue;
       if (SKIP_NAMES.has(binding.name)) continue;
       if (binding.name.startsWith("_")) continue;
+      if (binding.scope === "module" && (binding.kind === "const" || binding.kind === "function" || binding.kind === "class" || binding.kind === "type" || binding.kind === "interface" || binding.kind === "enum")) {
+        continue;
+      }
       issues.push({
         ruleId: "dead/unused-local",
         category: "logic",
@@ -36030,7 +36050,7 @@ var brokenLinkRule = createRule({
 
 // src/rules/dup/identical-block.ts
 import * as crypto from "crypto";
-var WINDOW_SIZE = 10;
+var WINDOW_SIZE = 20;
 var MIN_NORMALIZED_LENGTH = 40;
 var HASH_PREFIX_LENGTH = 16;
 var DEDUP_CACHE = /* @__PURE__ */ new Map();
@@ -36255,6 +36275,7 @@ var javaArraylistVsLinkedlistRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     let m;
     NEW_LINKED_LIST_REGEX.lastIndex = 0;
     while ((m = NEW_LINKED_LIST_REGEX.exec(source)) !== null) {
@@ -36289,6 +36310,7 @@ var javaEmptyCatchBlockRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     let m;
     SINGLE_LINE_EMPTY_CATCH_REGEX.lastIndex = 0;
     while ((m = SINGLE_LINE_EMPTY_CATCH_REGEX.exec(source)) !== null) {
@@ -36325,6 +36347,7 @@ var javaLegacyDateApiRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     const flagged = /* @__PURE__ */ new Set();
     let m;
     LEGACY_IMPORT_REGEX.lastIndex = 0;
@@ -36391,6 +36414,7 @@ var javaRawTypeOveruseRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     let m;
     RAW_TYPE_REGEX.lastIndex = 0;
     while ((m = RAW_TYPE_REGEX.exec(source)) !== null) {
@@ -36426,6 +36450,7 @@ var javaStringConcatLoopRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     if (!/\b(for|while|do)\b/.test(source)) return issues;
     let m;
     STRING_CONCAT_REGEX.lastIndex = 0;
@@ -36462,6 +36487,7 @@ var javaSystemOutPrintlnRule = createRule({
     const issues = [];
     const source = facts.v2?._source;
     if (!source) return issues;
+    if (!/\.java$/i.test(facts.filePath)) return issues;
     const matches = [];
     let m;
     PRINTLN_REGEX.lastIndex = 0;
@@ -41988,7 +42014,8 @@ var signal_strength_default = {
     precision: 0.6803,
     lastCalibratedAt: "2026-07-01T00:00:00Z",
     verdict: "USEFUL",
-    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=27874, FP=13097, P=68.0%, FPR=5.20%, lift=13.09. v7 was USEFUL (TP=16673, FP=10112, lift=11.29). v8 was USEFUL (TP=11201, FP=2985).",
+    defaultOff: true,
+    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=27874, FP=13097, P=68.0%, FPR=5.20%, lift=13.09. v7 was USEFUL (TP=16673, FP=10112, lift=11.29). v8 was USEFUL (TP=11201, FP=2985). v0.21.0: marked defaultOff because the v0.21.0 self-scan fires 109 times across 47 files \u2014 the heuristic counts `{`, `}`, `(`, `)` as raw characters without skipping string contents, template literals, regex literals, or comments. Files with normal closing-brace tails (the last 4-5 lines of a function definition) consistently fire because the tail has 0 opens and 3-4 closes. The corpus-level 13.09 lift doesn't translate to per-file discrimination. v0.22.0 should re-implement the count with a token-aware state machine (string/comment/regex skip) or switch to a parse-error-based signal.",
     aiSpecific: true,
     _v7Verdict: "USEFUL",
     _v7Lift: 11.29,
@@ -42091,7 +42118,8 @@ var signal_strength_default = {
     precision: 0.7495,
     lastCalibratedAt: "2026-07-01T00:00:00Z",
     verdict: "USEFUL",
-    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=59390, FP=19849, P=75.0%, FPR=7.87%, lift=9.52. v7 was USEFUL (TP=43498, FP=14983, lift=9.11). v8 was USEFUL (TP=15892, FP=4866).",
+    defaultOff: true,
+    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=59390, FP=19849, P=75.0%, FPR=7.87%, lift=9.52. v7 was USEFUL (TP=43498, FP=14983, lift=9.11). v8 was USEFUL (TP=15892, FP=4866). v0.21.0: marked defaultOff because the v0.21.0 self-scan fires 250 times across 250 files (1 per file) \u2014 the Binoculars-style per-segment entropy CV is too coarse a signal at the file level; the entropy floor is hit by most structured TypeScript regardless of whether the author was an LLM. The v8.5 corpus-level lift of 9.52 doesn't translate to per-file discrimination. v0.22.0 should tighten the threshold or re-calibrate against the per-file (not per-corpus) lift.",
     aiSpecific: true,
     _v7Verdict: "USEFUL",
     _v7Lift: 9.11,
@@ -42159,7 +42187,8 @@ var signal_strength_default = {
     precision: 0.5583,
     lastCalibratedAt: "2026-07-01T00:00:00Z",
     verdict: "USEFUL",
-    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=21567, FP=17061, P=55.8%, FPR=6.77%, lift=8.25. v7 was USEFUL (TP=17466, FP=13438, lift=7.71). v8 was USEFUL (TP=4101, FP=3623).",
+    defaultOff: true,
+    _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=21567, FP=17061, P=55.8%, FPR=6.77%, lift=8.25. v7 was USEFUL (TP=17466, FP=13438, lift=7.71). v8 was USEFUL (TP=4101, FP=3623). v0.21.2: marked defaultOff because precision 0.5583 < 0.60 floor (the v9.5 methodology in `docs/research/methodology-minimum-sample-size.md` sets the default-on threshold at P >= 0.60). At 55.8% precision, 44% of fires are false positives \u2014 too noisy to gate on. The rule stays available for opt-in via `rules: { 'ai/whitespace-regularity': 'low' }` in slopbrick.config.mjs. The 18 src/ self-scan fires are the typical noise profile.",
     aiSpecific: true,
     _v7Verdict: "USEFUL",
     _v7Lift: 7.71,
@@ -42507,7 +42536,8 @@ var signal_strength_default = {
     precision: 0,
     lastCalibratedAt: "2026-07-01T00:00:00Z",
     verdict: "DORMANT",
-    _calibrationNote: "v0.19: new rule, not yet calibrated. v8.5 calibration does not run dup/* rules. Scheduled for v0.20 calibration on near-dup corpus.",
+    defaultOff: false,
+    _calibrationNote: "v0.19: new rule, not yet calibrated. v8.5 calibration does not run dup/* rules (the calibration script analyzes each file in isolation, but dup/* rules require cross-file state \u2014 the module-scope dedup cache is per-process, not per-file, so per-file analysis never produces cross-file matches and the rule's recall reads as 0). Scheduled for v0.22 calibration on the v9 corpus with a per-project dedup cache. v0.21.1 self-scan: 207 fires in src/ after WINDOW_SIZE 10\u219220 (was 575) \u2014 69% reduction. The actual precision/recall of these 207 fires is unverified ground truth. Default stays on because the rule's value is real (block-level duplications are a classic AI-fingerprint) and the corpus-measured 0.0 is a methodology artifact, not a real measurement.",
     aiSpecific: false,
     _v7Verdict: "DORMANT",
     _v7Lift: 1,

package/dist/index.cjs

@@ -36,7 +36,7 @@ var VERSION;
 var init_header = __esm({
   "src/types/_header.ts"() {
     "use strict";
-    VERSION = "0.21.0";
+    VERSION = "0.21.2";
   }
 });
 
@@ -27003,6 +27003,16 @@ var init_segment_surprisal_cv = __esm({
       category: "ai",
       severity: "medium",
       aiSpecific: true,
+      // v0.21.0: defaultOff. The rule fires on 250/250 files in the
+      // self-scan (1 per file), which means it's not discriminating
+      // between AI and non-AI text — it fires on every file. The
+      // Binoculars-style per-segment entropy CV is too coarse a
+      // signal at the file level; the entropy floor is hit by most
+      // structured TypeScript regardless of whether the author was
+      // an LLM. Re-enable per-file via
+      // `rules: { 'ai/segment-surprisal-cv': 'medium' }` in
+      // slopbrick.config.mjs once a tighter threshold is calibrated.
+      defaultOff: true,
       description: "Per-segment cross-entropy CV is suspiciously low \u2014 Binoculars (Hans 2024): AI text has near-constant per-window entropy",
       create(context) {
         return context;
@@ -28295,6 +28305,7 @@ var init_unused_import = __esm({
           if (binding.kind !== "import-specifier" && binding.kind !== "import-default" && binding.kind !== "import-namespace") {
             continue;
           }
+          if (binding.isTypeOnly) continue;
           if (binding.isReferenced) continue;
           if (!binding.name) continue;
           const source = binding.source ? ` from '${binding.source}'` : "";
@@ -28361,6 +28372,9 @@ var init_unused_local = __esm({
           if (binding.isReferenced) continue;
           if (SKIP_NAMES.has(binding.name)) continue;
           if (binding.name.startsWith("_")) continue;
+          if (binding.scope === "module" && (binding.kind === "const" || binding.kind === "function" || binding.kind === "class" || binding.kind === "type" || binding.kind === "interface" || binding.kind === "enum")) {
+            continue;
+          }
           issues.push({
             ruleId: "dead/unused-local",
             category: "logic",
@@ -29596,7 +29610,7 @@ var init_identical_block = __esm({
     "use strict";
     crypto = __toESM(require("crypto"), 1);
     init_rule();
-    WINDOW_SIZE = 10;
+    WINDOW_SIZE = 20;
     MIN_NORMALIZED_LENGTH = 40;
     HASH_PREFIX_LENGTH = 16;
     DEDUP_CACHE = /* @__PURE__ */ new Map();
@@ -29842,6 +29856,7 @@ var init_arraylist_vs_linkedlist = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         let m;
         NEW_LINKED_LIST_REGEX.lastIndex = 0;
         while ((m = NEW_LINKED_LIST_REGEX.exec(source)) !== null) {
@@ -29883,6 +29898,7 @@ var init_empty_catch_block = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         let m;
         SINGLE_LINE_EMPTY_CATCH_REGEX.lastIndex = 0;
         while ((m = SINGLE_LINE_EMPTY_CATCH_REGEX.exec(source)) !== null) {
@@ -29926,6 +29942,7 @@ var init_legacy_date_api = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         const flagged = /* @__PURE__ */ new Set();
         let m;
         LEGACY_IMPORT_REGEX.lastIndex = 0;
@@ -29999,6 +30016,7 @@ var init_raw_type_overuse = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         let m;
         RAW_TYPE_REGEX.lastIndex = 0;
         while ((m = RAW_TYPE_REGEX.exec(source)) !== null) {
@@ -30041,6 +30059,7 @@ var init_string_concat_loop = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         if (!/\b(for|while|do)\b/.test(source)) return issues;
         let m;
         STRING_CONCAT_REGEX.lastIndex = 0;
@@ -30084,6 +30103,7 @@ var init_system_out_println = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         const matches = [];
         let m;
         PRINTLN_REGEX.lastIndex = 0;
@@ -43160,6 +43180,7 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
   if (source) {
     const { line, column } = positionFrom(node, vctx.lineOffsets);
     const importedNames = [];
+    const isTypeOnly = node.typeOnly === true;
     const specifiers = node.specifiers;
     if (Array.isArray(specifiers)) {
       for (const specifier of specifiers) {
@@ -43175,7 +43196,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           }
         } else if (specifier.type === "ImportSpecifier") {
@@ -43190,7 +43212,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           } else if (isObject(local) && local.type === "Identifier" && typeof local.value === "string") {
             const name = local.value;
@@ -43201,7 +43224,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           }
         }
@@ -43494,7 +43518,7 @@ function handleJSXOpeningElement(node, _parent, _path, vctx) {
   }
   return false;
 }
-function handleVariableDeclarator(node, _parent, path, vctx) {
+function handleVariableDeclarator(node, parent, path, vctx) {
   if (!isObject(node)) return false;
   const init = node.init;
   if (init && vctx.visit) {
@@ -43509,16 +43533,18 @@ function handleVariableDeclarator(node, _parent, path, vctx) {
     }
   }
   if (bindingNames.length > 0) {
-    const parent = node.parent;
-    const kind = isObject(parent) && parent.type === "VariableDeclaration" ? String(parent.kind ?? "var") : "var";
+    const declParent = isObject(parent) && parent.type === "ExportNamedDeclaration" ? parent.declaration ?? parent : parent;
+    const kind = isObject(declParent) && declParent.type === "VariableDeclaration" ? String(declParent.kind ?? "var") : "var";
     const { line, column } = positionFrom(id, vctx.lineOffsets);
+    const scope = vctx.ctx.stack.length === 0 ? "module" : "function";
     for (const bindingName of bindingNames) {
       vctx.facts.deadCode.bindings.push({
         name: bindingName,
         kind,
         line,
         column,
-        isReferenced: false
+        isReferenced: false,
+        scope
       });
     }
   }
@@ -47897,7 +47923,8 @@ var init_signal_strength = __esm({
         precision: 0.6803,
         lastCalibratedAt: "2026-07-01T00:00:00Z",
         verdict: "USEFUL",
-        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=27874, FP=13097, P=68.0%, FPR=5.20%, lift=13.09. v7 was USEFUL (TP=16673, FP=10112, lift=11.29). v8 was USEFUL (TP=11201, FP=2985).",
+        defaultOff: true,
+        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=27874, FP=13097, P=68.0%, FPR=5.20%, lift=13.09. v7 was USEFUL (TP=16673, FP=10112, lift=11.29). v8 was USEFUL (TP=11201, FP=2985). v0.21.0: marked defaultOff because the v0.21.0 self-scan fires 109 times across 47 files \u2014 the heuristic counts `{`, `}`, `(`, `)` as raw characters without skipping string contents, template literals, regex literals, or comments. Files with normal closing-brace tails (the last 4-5 lines of a function definition) consistently fire because the tail has 0 opens and 3-4 closes. The corpus-level 13.09 lift doesn't translate to per-file discrimination. v0.22.0 should re-implement the count with a token-aware state machine (string/comment/regex skip) or switch to a parse-error-based signal.",
         aiSpecific: true,
         _v7Verdict: "USEFUL",
         _v7Lift: 11.29,
@@ -48000,7 +48027,8 @@ var init_signal_strength = __esm({
         precision: 0.7495,
         lastCalibratedAt: "2026-07-01T00:00:00Z",
         verdict: "USEFUL",
-        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=59390, FP=19849, P=75.0%, FPR=7.87%, lift=9.52. v7 was USEFUL (TP=43498, FP=14983, lift=9.11). v8 was USEFUL (TP=15892, FP=4866).",
+        defaultOff: true,
+        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=59390, FP=19849, P=75.0%, FPR=7.87%, lift=9.52. v7 was USEFUL (TP=43498, FP=14983, lift=9.11). v8 was USEFUL (TP=15892, FP=4866). v0.21.0: marked defaultOff because the v0.21.0 self-scan fires 250 times across 250 files (1 per file) \u2014 the Binoculars-style per-segment entropy CV is too coarse a signal at the file level; the entropy floor is hit by most structured TypeScript regardless of whether the author was an LLM. The v8.5 corpus-level lift of 9.52 doesn't translate to per-file discrimination. v0.22.0 should tighten the threshold or re-calibrate against the per-file (not per-corpus) lift.",
         aiSpecific: true,
         _v7Verdict: "USEFUL",
         _v7Lift: 9.11,
@@ -48068,7 +48096,8 @@ var init_signal_strength = __esm({
         precision: 0.5583,
         lastCalibratedAt: "2026-07-01T00:00:00Z",
         verdict: "USEFUL",
-        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=21567, FP=17061, P=55.8%, FPR=6.77%, lift=8.25. v7 was USEFUL (TP=17466, FP=13438, lift=7.71). v8 was USEFUL (TP=4101, FP=3623).",
+        defaultOff: true,
+        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=21567, FP=17061, P=55.8%, FPR=6.77%, lift=8.25. v7 was USEFUL (TP=17466, FP=13438, lift=7.71). v8 was USEFUL (TP=4101, FP=3623). v0.21.2: marked defaultOff because precision 0.5583 < 0.60 floor (the v9.5 methodology in `docs/research/methodology-minimum-sample-size.md` sets the default-on threshold at P >= 0.60). At 55.8% precision, 44% of fires are false positives \u2014 too noisy to gate on. The rule stays available for opt-in via `rules: { 'ai/whitespace-regularity': 'low' }` in slopbrick.config.mjs. The 18 src/ self-scan fires are the typical noise profile.",
         aiSpecific: true,
         _v7Verdict: "USEFUL",
         _v7Lift: 7.71,
@@ -48416,7 +48445,8 @@ var init_signal_strength = __esm({
         precision: 0,
         lastCalibratedAt: "2026-07-01T00:00:00Z",
         verdict: "DORMANT",
-        _calibrationNote: "v0.19: new rule, not yet calibrated. v8.5 calibration does not run dup/* rules. Scheduled for v0.20 calibration on near-dup corpus.",
+        defaultOff: false,
+        _calibrationNote: "v0.19: new rule, not yet calibrated. v8.5 calibration does not run dup/* rules (the calibration script analyzes each file in isolation, but dup/* rules require cross-file state \u2014 the module-scope dedup cache is per-process, not per-file, so per-file analysis never produces cross-file matches and the rule's recall reads as 0). Scheduled for v0.22 calibration on the v9 corpus with a per-project dedup cache. v0.21.1 self-scan: 207 fires in src/ after WINDOW_SIZE 10\u219220 (was 575) \u2014 69% reduction. The actual precision/recall of these 207 fires is unverified ground truth. Default stays on because the rule's value is real (block-level duplications are a classic AI-fingerprint) and the corpus-measured 0.0 is a methodology artifact, not a real measurement.",
         aiSpecific: false,
         _v7Verdict: "DORMANT",
         _v7Lift: 1,
@@ -57063,7 +57093,6 @@ init_dist2();
 
 // src/cli/program.ts
 var import_node_path70 = require("path");
-var import_node_perf_hooks = require("perf_hooks");
 var import_commander2 = require("commander");
 
 // src/cli/options.ts
@@ -62212,7 +62241,7 @@ async function runCli({ start }) {
         await watchProject(options, cwd, paths);
         return;
       }
-      const scanStart = import_node_perf_hooks.performance.now();
+      const scanStart = performance.now();
       const {
         report,
         scores,
@@ -62222,8 +62251,8 @@ async function runCli({ start }) {
         baseline,
         machineReadableStdout
       } = await runScan(options, paths);
-      const scanElapsed = Math.round(import_node_perf_hooks.performance.now() - scanStart);
-      const totalElapsed = Math.round(import_node_perf_hooks.performance.now() - start);
+      const scanElapsed = Math.round(performance.now() - scanStart);
+      const totalElapsed = Math.round(performance.now() - start);
       if (options.baseline) {
         const cwd2 = (0, import_node_path70.resolve)(options.workspace ?? process.cwd());
         const configHash = hashConfig(config);

package/dist/index.d.cts

@@ -316,12 +316,29 @@ type BindingKind = 'import-specifier' | 'import-default' | 'import-namespace' |
  *  referenced-name set per file (intra-scope shadowing is rare enough
  *  in practice that file-level resolution catches ~95% of dead code
  *  with negligible false positives — the remaining 5% is mostly
- *  intentional re-exports which are handled separately). */
+ *  intentional re-exports which are handled separately).
+ *
+ *  `isTypeOnly` (v0.21.0) is set to true when the binding came from
+ *  an `import type { X }` declaration. The dead/unused-import rule
+ *  skips these unconditionally — a type-only import is always
+ *  "referenced" if the type is used (TypeScript guarantees it), and
+ *  TypeScript itself elides the import at build time, so the unused-
+ *  import rule has nothing actionable to report. */
 interface BindingRecord {
     /** The local name as written (`Button` for `import { Button } from ...`,
      *  `setFoo` for a parameter, etc.). */
     name: string;
     kind: BindingKind;
+    /** v0.21.0: true when this binding came from `import type { X }`. The
+     *  dead/unused-import rule skips these. */
+    isTypeOnly?: boolean;
+    /** v0.21.0: lexical scope of the binding. `'module'` for top-level
+     *  declarations outside any function; `'function'` for bindings
+     *  inside a function body. The dead/unused-local rule skips
+     *  `'module'` consts/functions/classes (they are often intentional
+     *  placeholder exports, type re-exports, or side-effect-ful
+     *  constructions — not real dead code). */
+    scope?: 'module' | 'function';
     /** Source line (1-indexed). */
     line: number;
     /** Source column (0-indexed). */

package/dist/index.d.ts

@@ -316,12 +316,29 @@ type BindingKind = 'import-specifier' | 'import-default' | 'import-namespace' |
  *  referenced-name set per file (intra-scope shadowing is rare enough
  *  in practice that file-level resolution catches ~95% of dead code
  *  with negligible false positives — the remaining 5% is mostly
- *  intentional re-exports which are handled separately). */
+ *  intentional re-exports which are handled separately).
+ *
+ *  `isTypeOnly` (v0.21.0) is set to true when the binding came from
+ *  an `import type { X }` declaration. The dead/unused-import rule
+ *  skips these unconditionally — a type-only import is always
+ *  "referenced" if the type is used (TypeScript guarantees it), and
+ *  TypeScript itself elides the import at build time, so the unused-
+ *  import rule has nothing actionable to report. */
 interface BindingRecord {
     /** The local name as written (`Button` for `import { Button } from ...`,
      *  `setFoo` for a parameter, etc.). */
     name: string;
     kind: BindingKind;
+    /** v0.21.0: true when this binding came from `import type { X }`. The
+     *  dead/unused-import rule skips these. */
+    isTypeOnly?: boolean;
+    /** v0.21.0: lexical scope of the binding. `'module'` for top-level
+     *  declarations outside any function; `'function'` for bindings
+     *  inside a function body. The dead/unused-local rule skips
+     *  `'module'` consts/functions/classes (they are often intentional
+     *  placeholder exports, type re-exports, or side-effect-ful
+     *  constructions — not real dead code). */
+    scope?: 'module' | 'function';
     /** Source line (1-indexed). */
     line: number;
     /** Source column (0-indexed). */

package/dist/index.js

@@ -19,7 +19,7 @@ var VERSION;
 var init_header = __esm({
   "src/types/_header.ts"() {
     "use strict";
-    VERSION = "0.21.0";
+    VERSION = "0.21.2";
   }
 });
 
@@ -26985,6 +26985,16 @@ var init_segment_surprisal_cv = __esm({
       category: "ai",
       severity: "medium",
       aiSpecific: true,
+      // v0.21.0: defaultOff. The rule fires on 250/250 files in the
+      // self-scan (1 per file), which means it's not discriminating
+      // between AI and non-AI text — it fires on every file. The
+      // Binoculars-style per-segment entropy CV is too coarse a
+      // signal at the file level; the entropy floor is hit by most
+      // structured TypeScript regardless of whether the author was
+      // an LLM. Re-enable per-file via
+      // `rules: { 'ai/segment-surprisal-cv': 'medium' }` in
+      // slopbrick.config.mjs once a tighter threshold is calibrated.
+      defaultOff: true,
       description: "Per-segment cross-entropy CV is suspiciously low \u2014 Binoculars (Hans 2024): AI text has near-constant per-window entropy",
       create(context) {
         return context;
@@ -28277,6 +28287,7 @@ var init_unused_import = __esm({
           if (binding.kind !== "import-specifier" && binding.kind !== "import-default" && binding.kind !== "import-namespace") {
             continue;
           }
+          if (binding.isTypeOnly) continue;
           if (binding.isReferenced) continue;
           if (!binding.name) continue;
           const source = binding.source ? ` from '${binding.source}'` : "";
@@ -28343,6 +28354,9 @@ var init_unused_local = __esm({
           if (binding.isReferenced) continue;
           if (SKIP_NAMES.has(binding.name)) continue;
           if (binding.name.startsWith("_")) continue;
+          if (binding.scope === "module" && (binding.kind === "const" || binding.kind === "function" || binding.kind === "class" || binding.kind === "type" || binding.kind === "interface" || binding.kind === "enum")) {
+            continue;
+          }
           issues.push({
             ruleId: "dead/unused-local",
             category: "logic",
@@ -29578,7 +29592,7 @@ var init_identical_block = __esm({
   "src/rules/dup/identical-block.ts"() {
     "use strict";
     init_rule();
-    WINDOW_SIZE = 10;
+    WINDOW_SIZE = 20;
     MIN_NORMALIZED_LENGTH = 40;
     HASH_PREFIX_LENGTH = 16;
     DEDUP_CACHE = /* @__PURE__ */ new Map();
@@ -29824,6 +29838,7 @@ var init_arraylist_vs_linkedlist = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         let m;
         NEW_LINKED_LIST_REGEX.lastIndex = 0;
         while ((m = NEW_LINKED_LIST_REGEX.exec(source)) !== null) {
@@ -29865,6 +29880,7 @@ var init_empty_catch_block = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         let m;
         SINGLE_LINE_EMPTY_CATCH_REGEX.lastIndex = 0;
         while ((m = SINGLE_LINE_EMPTY_CATCH_REGEX.exec(source)) !== null) {
@@ -29908,6 +29924,7 @@ var init_legacy_date_api = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         const flagged = /* @__PURE__ */ new Set();
         let m;
         LEGACY_IMPORT_REGEX.lastIndex = 0;
@@ -29981,6 +29998,7 @@ var init_raw_type_overuse = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         let m;
         RAW_TYPE_REGEX.lastIndex = 0;
         while ((m = RAW_TYPE_REGEX.exec(source)) !== null) {
@@ -30023,6 +30041,7 @@ var init_string_concat_loop = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         if (!/\b(for|while|do)\b/.test(source)) return issues;
         let m;
         STRING_CONCAT_REGEX.lastIndex = 0;
@@ -30066,6 +30085,7 @@ var init_system_out_println = __esm({
         const issues = [];
         const source = facts.v2?._source;
         if (!source) return issues;
+        if (!/\.java$/i.test(facts.filePath)) return issues;
         const matches = [];
         let m;
         PRINTLN_REGEX.lastIndex = 0;
@@ -43142,6 +43162,7 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
   if (source) {
     const { line, column } = positionFrom(node, vctx.lineOffsets);
     const importedNames = [];
+    const isTypeOnly = node.typeOnly === true;
     const specifiers = node.specifiers;
     if (Array.isArray(specifiers)) {
       for (const specifier of specifiers) {
@@ -43157,7 +43178,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           }
         } else if (specifier.type === "ImportSpecifier") {
@@ -43172,7 +43194,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           } else if (isObject(local) && local.type === "Identifier" && typeof local.value === "string") {
             const name = local.value;
@@ -43183,7 +43206,8 @@ function handleImportDeclaration(node, _parent, _path, vctx) {
               line,
               column,
               source,
-              isReferenced: false
+              isReferenced: false,
+              isTypeOnly
             });
           }
         }
@@ -43476,7 +43500,7 @@ function handleJSXOpeningElement(node, _parent, _path, vctx) {
   }
   return false;
 }
-function handleVariableDeclarator(node, _parent, path, vctx) {
+function handleVariableDeclarator(node, parent, path, vctx) {
   if (!isObject(node)) return false;
   const init = node.init;
   if (init && vctx.visit) {
@@ -43491,16 +43515,18 @@ function handleVariableDeclarator(node, _parent, path, vctx) {
     }
   }
   if (bindingNames.length > 0) {
-    const parent = node.parent;
-    const kind = isObject(parent) && parent.type === "VariableDeclaration" ? String(parent.kind ?? "var") : "var";
+    const declParent = isObject(parent) && parent.type === "ExportNamedDeclaration" ? parent.declaration ?? parent : parent;
+    const kind = isObject(declParent) && declParent.type === "VariableDeclaration" ? String(declParent.kind ?? "var") : "var";
     const { line, column } = positionFrom(id, vctx.lineOffsets);
+    const scope = vctx.ctx.stack.length === 0 ? "module" : "function";
     for (const bindingName of bindingNames) {
       vctx.facts.deadCode.bindings.push({
         name: bindingName,
         kind,
         line,
         column,
-        isReferenced: false
+        isReferenced: false,
+        scope
       });
     }
   }
@@ -47876,7 +47902,8 @@ var init_signal_strength = __esm({
         precision: 0.6803,
         lastCalibratedAt: "2026-07-01T00:00:00Z",
         verdict: "USEFUL",
-        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=27874, FP=13097, P=68.0%, FPR=5.20%, lift=13.09. v7 was USEFUL (TP=16673, FP=10112, lift=11.29). v8 was USEFUL (TP=11201, FP=2985).",
+        defaultOff: true,
+        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=27874, FP=13097, P=68.0%, FPR=5.20%, lift=13.09. v7 was USEFUL (TP=16673, FP=10112, lift=11.29). v8 was USEFUL (TP=11201, FP=2985). v0.21.0: marked defaultOff because the v0.21.0 self-scan fires 109 times across 47 files \u2014 the heuristic counts `{`, `}`, `(`, `)` as raw characters without skipping string contents, template literals, regex literals, or comments. Files with normal closing-brace tails (the last 4-5 lines of a function definition) consistently fire because the tail has 0 opens and 3-4 closes. The corpus-level 13.09 lift doesn't translate to per-file discrimination. v0.22.0 should re-implement the count with a token-aware state machine (string/comment/regex skip) or switch to a parse-error-based signal.",
         aiSpecific: true,
         _v7Verdict: "USEFUL",
         _v7Lift: 11.29,
@@ -47979,7 +48006,8 @@ var init_signal_strength = __esm({
         precision: 0.7495,
         lastCalibratedAt: "2026-07-01T00:00:00Z",
         verdict: "USEFUL",
-        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=59390, FP=19849, P=75.0%, FPR=7.87%, lift=9.52. v7 was USEFUL (TP=43498, FP=14983, lift=9.11). v8 was USEFUL (TP=15892, FP=4866).",
+        defaultOff: true,
+        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=59390, FP=19849, P=75.0%, FPR=7.87%, lift=9.52. v7 was USEFUL (TP=43498, FP=14983, lift=9.11). v8 was USEFUL (TP=15892, FP=4866). v0.21.0: marked defaultOff because the v0.21.0 self-scan fires 250 times across 250 files (1 per file) \u2014 the Binoculars-style per-segment entropy CV is too coarse a signal at the file level; the entropy floor is hit by most structured TypeScript regardless of whether the author was an LLM. The v8.5 corpus-level lift of 9.52 doesn't translate to per-file discrimination. v0.22.0 should tighten the threshold or re-calibrate against the per-file (not per-corpus) lift.",
         aiSpecific: true,
         _v7Verdict: "USEFUL",
         _v7Lift: 9.11,
@@ -48047,7 +48075,8 @@ var init_signal_strength = __esm({
         precision: 0.5583,
         lastCalibratedAt: "2026-07-01T00:00:00Z",
         verdict: "USEFUL",
-        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=21567, FP=17061, P=55.8%, FPR=6.77%, lift=8.25. v7 was USEFUL (TP=17466, FP=13438, lift=7.71). v8 was USEFUL (TP=4101, FP=3623).",
+        defaultOff: true,
+        _calibrationNote: "v8.5 calibration (v0.18.9, 2026-07-01): v7+v8 combined corpus (252080 neg + 294178 pos). v8.5 TP=21567, FP=17061, P=55.8%, FPR=6.77%, lift=8.25. v7 was USEFUL (TP=17466, FP=13438, lift=7.71). v8 was USEFUL (TP=4101, FP=3623). v0.21.2: marked defaultOff because precision 0.5583 < 0.60 floor (the v9.5 methodology in `docs/research/methodology-minimum-sample-size.md` sets the default-on threshold at P >= 0.60). At 55.8% precision, 44% of fires are false positives \u2014 too noisy to gate on. The rule stays available for opt-in via `rules: { 'ai/whitespace-regularity': 'low' }` in slopbrick.config.mjs. The 18 src/ self-scan fires are the typical noise profile.",
         aiSpecific: true,
         _v7Verdict: "USEFUL",
         _v7Lift: 7.71,
@@ -48395,7 +48424,8 @@ var init_signal_strength = __esm({
         precision: 0,
         lastCalibratedAt: "2026-07-01T00:00:00Z",
         verdict: "DORMANT",
-        _calibrationNote: "v0.19: new rule, not yet calibrated. v8.5 calibration does not run dup/* rules. Scheduled for v0.20 calibration on near-dup corpus.",
+        defaultOff: false,
+        _calibrationNote: "v0.19: new rule, not yet calibrated. v8.5 calibration does not run dup/* rules (the calibration script analyzes each file in isolation, but dup/* rules require cross-file state \u2014 the module-scope dedup cache is per-process, not per-file, so per-file analysis never produces cross-file matches and the rule's recall reads as 0). Scheduled for v0.22 calibration on the v9 corpus with a per-project dedup cache. v0.21.1 self-scan: 207 fires in src/ after WINDOW_SIZE 10\u219220 (was 575) \u2014 69% reduction. The actual precision/recall of these 207 fires is unverified ground truth. Default stays on because the rule's value is real (block-level duplications are a classic AI-fingerprint) and the corpus-measured 0.0 is a methodology artifact, not a real measurement.",
         aiSpecific: false,
         _v7Verdict: "DORMANT",
         _v7Lift: 1,
@@ -55926,10 +55956,7 @@ var init_renderOutput = __esm({
 });
 
 // src/cli/watch.ts
-import {
-  watch,
-  statSync as statSync7
-} from "fs";
+import { watch, statSync as statSync7 } from "fs";
 import { resolve as resolve11 } from "path";
 async function watchProject(options, cwd, paths) {
   let baselineMtime;
@@ -56966,7 +56993,6 @@ init_dist2();
 
 // src/cli/program.ts
 import { resolve as resolve41 } from "path";
-import { performance } from "perf_hooks";
 import { Command } from "commander";
 
 // src/cli/options.ts

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "slopbrick",
-  "version": "0.21.0",
-  "description": "Discovered, modeled, and governed repository structure. SlopBrick scans source code, classifies it against 95 rules in 15 categories, computes 4 scores (aiQuality, engineeringHygiene, security, repositoryHealth), and persists the structure for AI agents and CI.",
+  "version": "0.21.2",
+  "description": "Discovered, modeled, and governed repository structure. SlopBrick scans source code, classifies it against 95+ rules in 15 categories, computes 4 scores (aiSlopScore: lower=cleaner, engineeringHygiene, security, repositoryHealth composite), and persists the structure for AI agents and CI.",
   "type": "module",
   "bin": {
     "slopbrick": "bin/slopbrick.js"