fix: resolve OOM crashes from memory leaks in renderer process

[DeJeune]

What this PR does

Fixes #12925: Resolves three memory leak sources causing out-of-memory crashes during active agent/tool usage in the renderer process.

Before: Global maps and Redux caches accumulated entries indefinitely, causing heap growth until OOM termination.

After: Stream completion and topic switching now properly clean up resources, stabilizing heap memory.

Why we need it and why it was done in this way

Root causes identified:

1. globalActiveToolCalls (static Map): Entries added on tool-input-start but only removed on tool-result/error. Aborted streams left orphaned entries forever.
2. abortMap: Entries added per message but only removed on user abort. Normal stream completion never cleaned up entries.
3. Topic message caches: Redux messageIdsByTopic and entities grew unbounded as users viewed topics.

Fixes applied:

1. Added ToolCallChunkHandler.clearAll() static method, called in AiSdkToChunkAdapter.readFullStream() finally block.
2. Added cleanupAbortController(id) utility, called in both fetchAndProcessAgentResponseImpl and fetchAndProcessAssistantResponseImpl finally blocks.
3. Added evictTopicCache reducer and eviction logic in loadTopicMessagesThunk to keep only 5 most-recently-accessed topics in memory.

Why this approach:

• Minimal changes: Cleanups added to existing finally blocks (safe, always executes)
• No API changes: New functions are internal; no breaking changes
• Conservative cache limit: MAX_CACHED_TOPICS=5 balances memory vs user convenience (can adjust later)

Breaking changes

None. All changes are internal cleanup; no public APIs modified.

Checklist

• Lint, typecheck, format all pass
• All 3151 tests pass
• Changes follow existing patterns
• Code is simple and maintainable

[DeJeune]

Note

This comment was translated by Claude.

A simple fix that needs to be tested.

---

Original Content

a simple fix and need to be tested