Raise stale-costly defaults to 100k tokens / 60 min

Bump DEFAULT_STALE_COSTLY_THRESHOLD from 40k/5min to 100k/60min:

- tokens 100k: ~10% of the 1M context window, so the warning only fires
  when a cold prefix rebuild is a genuine cost (40k was ~$0.25 at Opus
  rates — too small to warrant a dialog).
- staleMs 60 min: the 5-minute value was the *default* ephemeral TTL,
  but the 1-hour opt-in TTL may be in effect (the Agent SDK decides, and
  we don't observe it). Using the longer bound means we only warn once
  the cache is almost certainly cold, avoiding false-positive nags while
  a still-warm cache makes continuation cheap.

Also corrects the misleading "5-minute TTL" doc comment.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RkLQSsdj7o4wQjcmjM1SD1

Author: claude

packages/core/src/sessions/contextUsage.ts

@@ -119,22 +119,31 @@ export interface StaleCostlyThreshold {
   /** Minimum context tokens for a conversation to count as "large". */
   tokens: number;
   /**
-   * Minimum idle time (ms) before a conversation counts as "stale". Set to
-   * roughly the Anthropic prompt-cache TTL: once the cache expires, the next
-   * turn re-processes the whole prefix at full input price instead of the
-   * ~10% cached-read rate.
+   * Minimum idle time (ms) before a conversation counts as "stale" — set at or
+   * above the prompt-cache TTL so that, once exceeded, the cache is expired and
+   * the next turn re-processes the whole prefix at full input price instead of
+   * the ~10% cached-read rate.
    */
   staleMs: number;
 }
 
 /**
- * Defaults for the stale-costly conversation warning: a conversation large
- * enough that a cold cache rebuild is noticeable, left idle past the default
- * 5-minute prompt-cache TTL.
+ * Defaults for the stale-costly conversation warning.
+ *
+ * `tokens` (100k): only conversations big enough that a cold prefix rebuild is
+ * a real cost — roughly 10% of the 1M context window — trip the warning.
+ *
+ * `staleMs` (60 min): Anthropic ephemeral caches default to a 5-minute TTL and
+ * can opt into a 1-hour one; the effective value depends on what the Agent SDK
+ * requests, which we don't control or observe. We deliberately use the longer
+ * 1-hour bound so the cache is almost certainly cold before we warn — warning
+ * while it is still warm would nag about a continuation that is in fact cheap,
+ * the worse failure. Erring long only means we occasionally skip a warning,
+ * never that we nag needlessly.
  */
 export const DEFAULT_STALE_COSTLY_THRESHOLD: StaleCostlyThreshold = {
-  tokens: 40_000,
-  staleMs: 5 * 60 * 1000,
+  tokens: 100_000,
+  staleMs: 60 * 60 * 1000,
 };
 
 /**