@adhd/backlog 0.1.3 → 0.1.4

package/package.json

@@ -1,24 +1,25 @@
 {
   "name": "@adhd/backlog",
-  "version": "0.1.3",
+  "version": "0.1.4",
   "bin": {
     "backlog": "index.js"
   },
   "dependencies": {
-    "@adhd/sox-graph-store": "^0.3.0",
-    "better-sqlite3": "^12.10.0",
+    "@adhd/apigen-core-client": "^0.3.0",
+    "@adhd/apigen-engine-naming": "^0.2.2",
+    "@adhd/apigen-plugin-api-fastify": "^0.2.3",
+    "@adhd/apigen-plugin-batch": "^0.2.2",
+    "@adhd/apigen-plugin-cli-output": "^0.2.3",
+    "@adhd/apigen-plugin-ir-cache": "^0.1.0",
+    "@adhd/apigen-plugin-mcp": "^0.2.3",
+    "@adhd/apigen-plugin-openapi": "^0.2.2",
     "@adhd/environment": "^0.1.3",
     "@adhd/environment-base-spec": "^0.1.0",
-    "@adhd/apigen-core-client": "^0.2.2",
-    "@adhd/apigen-plugin-api-fastify": "^0.2.2",
-    "@adhd/apigen-plugin-openapi": "^0.2.1",
-    "@adhd/apigen-plugin-mcp": "^0.2.2",
-    "@adhd/apigen-plugin-batch": "^0.2.1",
-    "@adhd/apigen-plugin-cli-output": "^0.2.2",
-    "@adhd/apigen-engine-naming": "^0.2.1",
-    "yaml": "1.10.3",
+    "@adhd/sox-graph-store": "^0.3.0",
+    "better-sqlite3": "^12.10.0",
     "pino": "10.3.1",
-    "pino-pretty": "13.1.3"
+    "pino-pretty": "13.1.3",
+    "yaml": "1.10.3"
   },
   "assets": [
     "skill"

package/store/crud.d.ts

@@ -1,6 +1,39 @@
 import { GraphBacklogStore } from './graph-backlog-store.js';
 import { BacklogItem, CreateItemInput, CreateItemResult, UpdateItemInput } from '../model.js';
 
+/**
+ * Minimum fraction of the NEW item's meaningful title tokens that must also
+ * appear in a CANDIDATE's title for the candidate to count as a duplicate
+ * (BUG-BACKLOG-DEDUPE-FTS-WEAK-MATCH-001). Exported so callers/tests can
+ * tune it without touching this file's internals.
+ *
+ * Why title-to-title overlap instead of the FTS hit's bm25 `score`:
+ * `store.graph.searchNodes()` (`@adhd/sox-graph-store`) DOES return a
+ * per-hit `score` (`-fts_node.rank`, i.e. a positive, higher-is-better bm25
+ * score) — but bm25 is corpus- and document-length-relative, not an absolute
+ * similarity measure, and empirically it does NOT discriminate this bug's
+ * failure mode: a live probe reproducing the exact reported case (a short
+ * generic new title vs. a long, verbose, unrelated document that happens to
+ * repeat a couple of the new title's common words many times) scored
+ * `~0.0000109`, while a genuine short-title-vs-short-title near-duplicate
+ * scored `~0.0000113` — the SAME order of magnitude, with no clean
+ * separating threshold between them. A long document's sheer token count
+ * inflates its bm25 term-frequency component enough to rival a real
+ * duplicate's score. Title-to-title token overlap has no such document-
+ * length confound: a new item's title can only ever match against a
+ * candidate's (typically similarly short) title, so incidental repetition
+ * buried in a long candidate BODY can no longer count toward "looks like a
+ * duplicate" at all.
+ *
+ * 0.5 (at least half the new title's meaningful tokens must recur in the
+ * candidate's title) was chosen because DESIGN.md §2.4's own worked example
+ * ("same bug, different words") is a near-total title rewrite that still
+ * shares most of its content words ("database connection pool leaks under
+ * load" -> "the database connection pool leaks under load" is 6/6); dropping
+ * to a small minority match (e.g. 1-2 shared generic words out of 5) is
+ * exactly the false-positive shape this bug reports and must NOT pass.
+ */
+export declare const TITLE_OVERLAP_MIN_FRACTION = 0.5;
 declare function dedupeScan(store: GraphBacklogStore, repo: string, input: CreateItemInput): BacklogItem[];
 export declare function createItemNode(store: GraphBacklogStore, input: CreateItemInput): CreateItemResult;
 export declare function getItemNode(store: GraphBacklogStore, repo: string, humanId: string): BacklogItem | null;