#4999: fix(memory-flush): use contextTokens instead of totalTokens for threshold check

## Problem `shouldRunMemoryFlush()` was using `totalTokens` from the session entry to decide if a memory flush should trigger before compaction. However, `totalTokens` is set by `session-usage.ts` with **tokens from the last turn only**, not the cumulative context size. This caused memory flush to never trigger because it compared ~30k tokens (last turn) against the ~176k threshold, even when actual context was 188k tokens. ## Solution Changed `shouldRunMemoryFlush()` to use `contextTokens` instead, which correctly represents the actual context window usage. ## Changes - `src/auto-reply/reply/memory-flush.ts`: Use `contextTokens` instead of `totalTokens` - `src/auto-reply/reply/memory-flush.test.ts`: Updated tests to match <!-- greptile_comment --> <h2>Greptile Overview</h2> <h3>Greptile Summary</h3> This PR fixes pre-compaction memory flush triggering by switching the threshold check in `shouldRunMemoryFlush()` from `totalTokens` (apparently last-turn tokens) to `contextTokens` (actual context window usage), and updates the associated unit tests accordingly. Within the auto-reply compaction/memory-flush flow, this makes flush decisions reflect the real prompt/context size so the flush can trigger when the session is near the context limit. <h3>Confidence Score: 3/5</h3> - Mostly safe to merge, but includes an accidental symlink that should be removed. - The core logic change is small and well-covered by updated tests, but the PR also introduces a `workspace` symlink to an absolute local path which will break in other environments and is likely unintended. - workspace <!-- greptile_other_comments_section --> _{(2/5) Greptile learns from your feedback when you react with thumbs up/down!} <!-- /greptile_comment -->