fix: treat zero-reset Antigravity 429s as transient (#8626)

Validated in local merge-train T7 (ungrouped batch 2)

open-sse/services/antigravity429Engine.ts

@@ -61,6 +61,14 @@ const FULL_QUOTA_COOLDOWN_MS = 24 * 60 * 60 * 1000; // 24 hours
 export function classify429(errorMessage: string): Category {
   const lower = (errorMessage || "").toLowerCase();
 
+  // Cloud Code may report an exhausted-capacity message with a zero reset
+  // window for a burst/RPM throttle. The explicit zero reset is stronger
+  // evidence than the generic wording, so retry briefly instead of applying
+  // the durable quota cooldown.
+  if (/\breset\s+(?:after|in)\s+0s\b/.test(lower)) {
+    return "rate_limited";
+  }
+
   // Check for quota exhaustion first (most specific)
   for (const kw of QUOTA_EXHAUSTED_KEYWORDS) {
     if (lower.includes(kw)) return "quota_exhausted";

tests/unit/antigravity-429-quota-cooldown.test.ts

@@ -70,6 +70,16 @@ test("classify429: standard Gemini rate limit 'resource has been exhausted' -> r
   );
 });
 
+test("classify429: exhausted capacity with reset after 0s is rate_limited", () => {
+  const message = "You have exhausted your capacity on this model. Your quota will reset after 0s.";
+  const category = classify429(message);
+  assert.equal(category, "rate_limited");
+
+  const decision = decide429(category, 2_000);
+  assert.equal(decision.kind, "soft_retry");
+  assert.equal(decision.retryAfterMs, 2_000);
+});
+
 // ── DB persistence (the missing wire — Bug #2) ───────────────────────────────
 
 test("markConnectionQuotaExhausted persists 24h cooldown; isConnectionRateLimited returns true", async () => {