Greptile Summary

This PR fixes a bug in runMemoryFlushIfNeeded where memoryFlushCompactionCount was unconditionally persisted on every flush turn, even when compaction didn't actually complete during that turn. This caused the counters compactionCount and memoryFlushCompactionCount to synchronize, making hasAlreadyFlushedForCurrentCompaction return true on the next cycle and producing the "flush, skip, flush, skip" alternating pattern.

The fix is a minimal, targeted 4-line conditional spread: memoryFlushCompactionCount is now only written to the session store when memoryCompactionCompleted is true. This keeps the counters desynced when compaction doesn't run, ensuring shouldRunMemoryFlush continues to return true on subsequent cycles as intended.

• Core change: ...(memoryCompactionCompleted ? { memoryFlushCompactionCount } : {}) replaces the unconditional memoryFlushCompactionCount in the updateSessionStoreEntry call
• Side effect: memoryFlushAt is still always updated (even without compaction), which is correct for tracking flush timestamp independently of the compaction counter
• Backward compatibility: Existing sessions with synced counters will naturally desync on the next compaction cycle, as noted in the PR description

Confidence Score: 5/5

• This PR is safe to merge — it is a small, well-scoped bug fix with no side effects on unrelated code paths.
• The change is minimal (4 lines, one file), directly addresses the described root cause, and the logic is verified against hasAlreadyFlushedForCurrentCompaction and shouldRunMemoryFlush in memory-flush.ts. The dedup guard still prevents double-flushing within a compaction cycle. The memoryFlushAt timestamp continues to be written unconditionally, which is correct. No regressions introduced.
• No files require special attention.

_{Last reviewed commit: 7e5fb5d}

Closing to manage active PR count. Will reopen when slot is available.