LaTeX-escape restoration at the display seam (owner ruling 2026-09-26): JSON-escape debris (FF ate \forall's f, VT \varphi's v, TAB \to's t) restored inside balanced math spans — one seam (tool-parts parse) heals new turns and poisoned records; outside spans untouched

2026-09-27 02:19:27 +03:00
parent d6d7c09e2e
commit 8faf97edc7
3 changed files with 137 additions and 3 deletions
@@ -214,3 +214,56 @@ export function transformMathDelimiters(markdown: string): string {

 	return out.join("");
 }
+
+/**
+ * LaTeX-escape restoration (owner ruling 2026-09-26) — the JSON-debris fix.
+ *
+ * When a model emits LaTeX with SINGLE backslashes inside JSON tool-call
+ * arguments (`\forall`, not the properly escaped `\\forall`), JSON escape
+ * resolution eats the command's first letter: `\f` → form feed + "orall",
+ * `\v` → vertical tab + "arphi", `\t` → tab + "o". The debris survives into
+ * the persisted RAW args and renders as garbage. Inside a MATH SPAN these
+ * control chars can never be legitimate content, so restoring the two-char
+ * escape reconstructs the model's intent byte-perfectly. Outside spans the
+ * same chars might be real prose (pasted tables, formatted text), so only
+ * balanced `\(…\)` / `\[…\]` spans are touched — the same scanner rules as
+ * transformMathDelimiters (escaped delimiters are literal, unbalanced
+ * delimiters pass through untouched).
+ */
+const ESCAPE_DEBRIS: Record<string, string> = {
+	"\b": "\\b",
+	"\t": "\\t",
+	"\n": "\\n",
+	"\v": "\\v",
+	"\f": "\\f",
+	"\r": "\\r",
+};
+
+const DEBRIS_RE = /[\b\t\n\v\f\r]/;
+
+export function restoreMathEscapes(text: string): string {
+	if (!DEBRIS_RE.test(text)) return text;
+	const out: string[] = [];
+	let i = 0;
+	while (i < text.length) {
+		const ch = text[i]!;
+		if (ch === "\\" && !isEscaped(text, i)) {
+			const next = text[i + 1];
+			if (next === "(" || next === "[") {
+				const isDisplay = next === "[";
+				const open = isDisplay ? DISPLAY_OPEN : INLINE_OPEN;
+				const close = isDisplay ? DISPLAY_CLOSE : INLINE_CLOSE;
+				const closeAt = findClose(text, i + open.length, close);
+				if (closeAt !== -1) {
+					const inner = text.slice(i + open.length, closeAt);
+					out.push(open + inner.replace(DEBRIS_RE, (c) => ESCAPE_DEBRIS[c]!) + close);
+					i = closeAt + close.length;
+					continue;
+				}
+			}
+		}
+		out.push(ch);
+		i += 1;
+	}
+	return out.join("");
+}
@@ -30,6 +30,7 @@
 import { Value } from "typebox/value";
 import type { TSchema } from "typebox";
 import { quizInputBase, askInputSchema, askOptionProblems, quizInputProblems, type AskToolInput, type QuizToolInput } from "./question-core.ts";
+import { restoreMathEscapes } from "./math-delims.ts";
 import { probeDoneInputSchema, type ProbeDoneInput } from "./contracts/probe.ts";

 /** The pi-boundary parse step, mirrored here (live-smoke fix, 2026-09-26):
@@ -65,8 +66,25 @@ function boundaryParse<T>(schema: TSchema, input: unknown): T | undefined {
 * never have been posed (e.g. `correctAnswer: undefined`, which Convert
 * widens to "") renders as a quiet fallback, not a broken quiz.
 */
+/** LaTeX-escape restoration, deep over every string field (the JSON-debris
+ *  fix, owner ruling 2026-09-26): records persist RAW call args, so
+ *  single-backslash LaTeX eaten by JSON escape resolution (FF ate the 'f' of
+ *  \forall, VT the 'v' of \varphi, TAB the 't' of \to) survives to the
+ *  display path. This is the ONE seam every rendered question passes
+ *  through — new turns and already-poisoned records alike heal here. */
+function restoreMathEscapesDeep(value: unknown): unknown {
+	if (typeof value === "string") return restoreMathEscapes(value);
+	if (Array.isArray(value)) return value.map(restoreMathEscapesDeep);
+	if (typeof value === "object" && value !== null) {
+		const out: Record<string, unknown> = {};
+		for (const [k, v] of Object.entries(value)) out[k] = restoreMathEscapesDeep(v);
+		return out;
+	}
+	return value;
+}
+
 export function parseQuizInput(input: unknown): QuizToolInput | undefined {
-	const cleaned = boundaryParse<QuizToolInput>(quizInputBase, input);
+	const cleaned = boundaryParse<QuizToolInput>(quizInputBase, restoreMathEscapesDeep(input));
 	if (!cleaned) return undefined;
 	if (quizInputProblems(cleaned).length > 0) return undefined;
 	return cleaned;
@@ -79,7 +97,7 @@ export function parseQuizInput(input: unknown): QuizToolInput | undefined {
 * keying, so they parse to undefined exactly like the zod seam did).
 */
 export function parseAskInput(input: unknown): AskToolInput | undefined {
-	const cleaned = boundaryParse<AskToolInput>(askInputSchema, input);
+	const cleaned = boundaryParse<AskToolInput>(askInputSchema, restoreMathEscapesDeep(input));
 	if (!cleaned) return undefined;
 	if (askOptionProblems(cleaned.options ?? []).length > 0) return undefined;
 	return cleaned;
@@ -152,4 +152,67 @@ describe("parseProbeDoneInput", () => {
 	it("a mid-stream partial (no strands yet) does not parse", () => {
 		assert.equal(parseProbeDoneInput({ assessment: { summary: "…" } }), undefined);
 	});
-});
+});
+describe("LaTeX-escape restoration (the JSON-debris fix, owner ruling 2026-09-26)", () => {
+	// The exact live shapes from the modal-logic session: single-backslash
+	// LaTeX inside JSON args — JSON escape resolution ate the command's
+	// first letter (\f → FF ate 'f' of \forall, \v → VT ate 'v' of \varphi).
+	it("parseQuizInput restores escape debris inside math spans (question + options + explanation)", () => {
+		const parsed = parseQuizInput({
+			question: "What does \\(\\forall x\\, \\varphi(x)\\) mean?",
+			options: [
+				{ label: "\\(\x0barphi\\) is true when every element satisfies it", value: "tarski" },
+				{ label: "\\(\x0corall\\) and \\(\\exists\\) are just shorthand", value: "shorthand" },
+			],
+			correctAnswer: "tarski",
+			explanation: "\\(\x0corall\\) quantifies over the domain.",
+		});
+		assert.ok(parsed);
+		assert.equal(parsed.question, "What does \\(\\forall x\\, \\varphi(x)\\) mean?");
+		assert.equal(parsed.options[0]!.label, "\\(\\varphi\\) is true when every element satisfies it");
+		assert.equal(parsed.options[1]!.label, "\\(\\forall\\) and \\(\\exists\\) are just shorthand");
+		assert.equal(parsed.explanation, "\\(\\forall\\) quantifies over the domain.");
+	});
+
+	it("parseAskInput restores debris in ask option labels too", () => {
+		const parsed = parseAskInput({
+			question: "Goal?",
+			options: [{ label: "\\(\x0corall\\) rigor" }, { label: "Intuition first" }],
+		});
+		assert.ok(parsed);
+		assert.equal(parsed.options![0]!.label, "\\(\\forall\\) rigor");
+	});
+
+	it("restores TAB debris (the \\to case) inside spans only", () => {
+		const parsed = parseQuizInput({
+			question: "Is \\(p\to q\\) valid?",
+			options: [{ label: "yes" }, { label: "no" }],
+			correctAnswer: "no",
+			explanation: "",
+		});
+		assert.ok(parsed);
+		assert.equal(parsed.question, "Is \\(p\\to q\\) valid?");
+	});
+
+	it("control chars OUTSIDE math spans are left alone (real prose/pasted text)", () => {
+		const parsed = parseQuizInput({
+			question: "Pick:\ta\x0bb",
+			options: [{ label: "x" }, { label: "y" }],
+			correctAnswer: "x",
+			explanation: "",
+		});
+		assert.ok(parsed);
+		assert.equal(parsed.question, "Pick:\ta\x0bb");
+	});
+
+	it("unbalanced delimiters pass through untouched (mid-stream safety)", () => {
+		const parsed = parseQuizInput({
+			question: "\\(\\x0corall",
+			options: [{ label: "x" }, { label: "y" }],
+			correctAnswer: "x",
+			explanation: "",
+		});
+		assert.ok(parsed);
+		assert.equal(parsed.question, "\\(\\x0corall");
+	});
+});