#8900: fix: improve default memory flush prompts with urgency and specifics

## Problem The default memory flush prompts before compaction are too vague: **Before:** ``` Pre-compaction memory flush. Store durable memories now (use memory/YYYY-MM-DD.md; create memory/ if needed). If nothing to store, reply with NO_REPLY. ``` This doesn't: - Communicate urgency - Specify what to capture - Explain what will be lost - Motivate the model to actually do it ## Solution New prompts with: - Visual urgency indicators (⚠️, 🚨) - Specific checklist of what to capture - Clear explanation of consequences - Warning that summarizer won't preserve active work **After:** ``` ⚠️ COMPACTION IMMINENT - Save your work NOW. Write to memory/YYYY-MM-DD.md (create memory/ if needed): 1. ACTIVE TASKS: What are you currently working on? Status of each. 2. RECENT REQUESTS: Last 2-3 things the user asked for. 3. IN-PROGRESS WORK: Any batch operations, their progress. 4. BLOCKERS: Anything stuck or waiting. 5. PENDING COMMITMENTS: Promises made that need follow-up. This context will be LOST after compaction. The summary alone won't preserve task state. ``` ## Why This Matters Models treat vague instructions as optional. When memory flush says 'store durable memories' without specifics, the model often replies NO_REPLY because it doesn't know what's durable. By being explicit about what to save and why, we increase the likelihood that active work survives compaction. <!-- greptile_comment --> <h2>Greptile Overview</h2> <h3>Greptile Summary</h3> This PR updates the default pre-compaction memory flush prompts in `src/auto-reply/reply/memory-flush.ts` to be more urgent and specific, switching from a single-line, generic instruction to multi-line prompts with an explicit checklist of what to persist (tasks, recent requests, batch progress, blockers, commitments) and clearer consequences if not saved. The behavior remains configuration-driven via `resolveMemoryFlushSettings`, with the prompts still passed through `ensureNoReplyHint()` to guarantee the silent-reply token guidance is present. <h3>Confidence Score: 4/5</h3> - This PR is low risk and mainly updates prompt strings. - Changes are confined to string constants used for memory flush prompts; no control flow or data handling changed. The main potential downside is UX/compatibility around emoji rendering and slightly mixed instruction wording, but there are no obvious runtime or type-safety risks. - src/auto-reply/reply/memory-flush.ts <!-- greptile_other_comments_section --> **Context used:** - Context from `dashboard` - CLAUDE.md ([source](https://app.greptile.com/review/custom-context?memory=fd949e91-5c3a-4ab5-90a1-cbe184fd6ce8)) - Context from `dashboard` - AGENTS.md ([source](https://app.greptile.com/review/custom-context?memory=0d0c8278-ef8e-4d6c-ab21-f5527e322f13)) <!-- /greptile_comment -->