fix(mnemopi): dropped object-shaped fact coercion

- Unwrapped extractor fact items from known text fields instead of coercing arbitrary objects with String().

- Dropped unrecognized object entries so derived fact indexes cannot persist literal [object Object] rows.

- Added a parseFacts regression covering mixed strings, wrapped objects, and junk timeline objects.

Fixes #4649

packages/mnemopi/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed extractor JSON parsing so object-shaped fact, instruction, preference, and timeline items are unwrapped from known text fields or dropped instead of persisting literal `[object Object]` rows. ([#4649](https://github.com/can1357/oh-my-pi/issues/4649))
+
 ## [16.3.7] - 2026-07-05
 
 ### Added

packages/mnemopi/src/core/extraction.ts

@@ -85,6 +85,7 @@ function stripFence(raw: string): string {
 const FLAT_FACT_LIMIT = 5;
 const STRUCTURED_CATEGORY_LIMIT = 5;
 const STRING_CATEGORY_KEYS = ["facts", "instructions", "preferences", "timelines"] as const;
+const FACT_TEXT_FIELD_KEYS = ["fact", "text", "content", "value", "statement"] as const;
 
 /** Parsed knowledge-graph edge emitted by the extractor LLM. */
 export interface ExtractedKgTriple {
@@ -118,8 +119,20 @@ function normalizeFactArray(items: unknown): string[] {
 	}
 	const out: string[] = [];
 	for (const item of items) {
-		if (item !== null && item !== undefined && String(item).trim() !== "") {
-			const normalized = normalizeFact(String(item));
+		let text: string | null = null;
+		if (typeof item === "string") {
+			text = item.trim();
+		} else if (isRecord(item)) {
+			for (const key of FACT_TEXT_FIELD_KEYS) {
+				const candidate = item[key];
+				if (typeof candidate === "string" && candidate.trim() !== "") {
+					text = candidate.trim();
+					break;
+				}
+			}
+		}
+		if (text !== null && text !== "") {
+			const normalized = normalizeFact(text);
 			if (normalized !== "") {
 				out.push(normalized);
 				if (out.length >= STRUCTURED_CATEGORY_LIMIT) break;

packages/mnemopi/test/extraction.test.ts

@@ -52,6 +52,22 @@ describe("structured extraction", () => {
 		expect(parseFacts("NO_FACTS")).toEqual([]);
 	});
 
+	it("drops unrecognized object facts instead of stringifying them", () => {
+		const modelJson = JSON.stringify({
+			facts: [{ fact: "The user prefers tabs over spaces" }, { nested: {} }, "The user likes concise replies."],
+			instructions: [{ text: "Always include verification details" }],
+			preferences: [{ value: "Prefers dark mode" }],
+			timelines: [{ subject: "release", predicate: "on", object: "2026-08-01" }],
+		});
+
+		expect(parseFacts(modelJson)).toEqual([
+			"The user prefers tabs over spaces",
+			"The user likes concise replies",
+			"Always include verification details",
+			"Prefers dark mode",
+		]);
+	});
+
 	it("treats a valid empty structured extraction as no facts", () => {
 		expect(parseFacts('{"facts": [], "instructions": [], "preferences": [], "timelines": [], "kg": []}')).toEqual([]);
 		expect(