pi hygiene-walk perf test: assert the memo ratio (load-invariant) instead of a wall-clock budget; absolute ceiling behind MC_PERF_GATE

Co-authored-by: Alfonso <alfonso-magic-context@users.noreply.github.com>

Author: ualtinok

packages/pi-plugin/src/tail-hygiene-walk-pi.test.ts

@@ -840,26 +840,49 @@ describe("TS/Pi/module differential hygiene corpus", () => {
 });
 
 describe("Pi hygiene walk performance", () => {
-	it("stays below 15ms p95 on a memoized 250k-token rendered tail", () => {
+	it("memoized 250k-token rendered tail walks are cheap relative to the cold walk", () => {
 		const messages = [textMessage("user", "token ".repeat(250_000))];
 		const tags = [tag(1, "perf:p0", "message")];
 		const stableId = withStableIds(messages, ["perf"]);
-		const durations: number[] = [];
-		for (let iteration = 0; iteration < 25; iteration += 1) {
+		const walk = (content: string) => {
+			(messages[0] as { content: string }).content = content;
 			const start = performance.now();
 			measurePiTailHygiene({
 				messages,
 				tags,
 				protectedTagNumbers: new Set(),
 				stableId,
 			});
-			durations.push(performance.now() - start);
+			return performance.now() - start;
+		};
+		const p95Of = (samples: number[]) => {
+			const sorted = [...samples].sort((left, right) => left - right);
+			return (
+				sorted[Math.ceil(sorted.length * 0.95) - 1] ?? Number.POSITIVE_INFINITY
+			);
+		};
+		// The content memo is keyed on the rendered text, so a walk over unchanged
+		// content must skip tokenization while a walk over fresh content pays it.
+		// A shared CI runner cannot promise an absolute millisecond budget (the
+		// memoized walk read 2.7ms locally and 19ms on a loaded runner), so the
+		// invariant is the ratio between the unmemoized and memoized walks measured
+		// in the same process, which load scales equally. The absolute ceiling is
+		// kept behind MC_PERF_GATE for machines that opt into wall-clock budgets.
+		const base = "token ".repeat(250_000);
+		const unmemoized: number[] = [];
+		for (let iteration = 0; iteration < 8; iteration += 1) {
+			unmemoized.push(walk(`${base} fresh-${iteration}`));
 		}
-		durations.sort((left, right) => left - right);
-		const p95 =
-			durations[Math.ceil(durations.length * 0.95) - 1] ??
-			Number.POSITIVE_INFINITY;
-		console.log(`pi-tail-hygiene-walk 250k-token p95=${p95.toFixed(3)}ms`);
-		expect(p95).toBeLessThan(15);
+		walk(base);
+		const memoized: number[] = [];
+		for (let iteration = 0; iteration < 25; iteration += 1)
+			memoized.push(walk(base));
+		const unmemoizedP95 = p95Of(unmemoized);
+		const p95 = p95Of(memoized);
+		console.log(
+			`pi-tail-hygiene-walk 250k-token unmemoized p95=${unmemoizedP95.toFixed(3)}ms memoized p95=${p95.toFixed(3)}ms`,
+		);
+		expect(p95).toBeLessThan(unmemoizedP95 / 5);
+		if (process.env.MC_PERF_GATE === "1") expect(p95).toBeLessThan(15);
 	});
 });