fix: stop defaulting to rate_limit when cooldown reason is unknown

When a profile enters cooldown but no failure counts are recorded (e.g.
a transient network error or unclassified server error), the inferred
reason was hardcoded to "rate_limit". This caused misleading "API rate
limit reached" warnings for users who hadn't actually hit any rate limit.

Change both default-reason sites to use "unknown" instead:
- resolveProfilesUnavailableReason fallback in usage.ts
- resolveCooldownDecision null-coalesce in model-fallback.ts

Treat "unknown" as a transient reason for probe and same-provider
fallback purposes, matching the existing behavior for rate_limit and
overloaded cooldowns so recovery attempts still work.

Fixes #32828

src/agents/auth-profiles/usage.test.ts

@@ -207,7 +207,7 @@ describe("resolveProfilesUnavailableReason", () => {
     ).toBe("overloaded");
   });
 
-  it("falls back to rate_limit when active cooldown has no reason history", () => {
+  it("falls back to unknown when active cooldown has no reason history", () => {
     const now = Date.now();
     const store = makeStore({
       "anthropic:default": {
@@ -221,7 +221,7 @@ describe("resolveProfilesUnavailableReason", () => {
         profileIds: ["anthropic:default"],
         now,
       }),
-    ).toBe("rate_limit");
+    ).toBe("unknown");
   });
 
   it("ignores expired windows and returns null when no profile is actively unavailable", () => {

src/agents/auth-profiles/usage.ts

@@ -110,7 +110,11 @@ export function resolveProfilesUnavailableReason(params: {
       recordedReason = true;
     }
     if (!recordedReason) {
-      addScore("rate_limit", 1);
+      // No failure counts recorded for this cooldown window. Previously this
+      // defaulted to "rate_limit", which caused false "rate limit reached"
+      // warnings when the actual reason was unknown (e.g. transient network
+      // blip or server error without a classified failure count).
+      addScore("unknown", 1);
     }
   }
 

src/agents/model-fallback.ts

@@ -449,7 +449,7 @@ function resolveCooldownDecision(params: {
       store: params.authStore,
       profileIds: params.profileIds,
       now: params.now,
-    }) ?? "rate_limit";
+    }) ?? "unknown";
   const isPersistentAuthIssue = inferredReason === "auth" || inferredReason === "auth_permanent";
   if (isPersistentAuthIssue) {
     return {
@@ -483,7 +483,10 @@ function resolveCooldownDecision(params: {
   // limits, which are often model-scoped and can recover on a sibling model.
   const shouldAttemptDespiteCooldown =
     (params.isPrimary && (!params.requestedModel || shouldProbe)) ||
-    (!params.isPrimary && (inferredReason === "rate_limit" || inferredReason === "overloaded"));
+    (!params.isPrimary &&
+      (inferredReason === "rate_limit" ||
+        inferredReason === "overloaded" ||
+        inferredReason === "unknown"));
   if (!shouldAttemptDespiteCooldown) {
     return {
       type: "skip",
@@ -588,13 +591,16 @@ export async function runWithModelFallback<T>(params: {
         if (
           decision.reason === "rate_limit" ||
           decision.reason === "overloaded" ||
-          decision.reason === "billing"
+          decision.reason === "billing" ||
+          decision.reason === "unknown"
         ) {
           // Probe at most once per provider per fallback run when all profiles
           // are cooldowned. Re-probing every same-provider candidate can stall
           // cross-provider fallback on providers with long internal retries.
           const isTransientCooldownReason =
-            decision.reason === "rate_limit" || decision.reason === "overloaded";
+            decision.reason === "rate_limit" ||
+            decision.reason === "overloaded" ||
+            decision.reason === "unknown";
           if (isTransientCooldownReason && cooldownProbeUsedProviders.has(candidate.provider)) {
             const error = `Provider ${candidate.provider} is in cooldown (probe already attempted this run)`;
             attempts.push({