[FEATURE] Allow custom instructions for auto-compact via settings or CLAUDE.md

Problem Statement

When auto-compact triggers, the generated summaries are often too generic and lose critical details like:

• Specific file changes and paths
• Error messages and debugging steps
• Architectural decisions and their rationale
• Function names, configuration values, and implementation details

This forces users to redo work after auto-compact because Claude loses context about what was already accomplished. The only current workaround is disabling auto-compact entirely and managing context manually.

Proposed Solution

Allow users to define custom instructions for auto-compact summaries via:

1. settings.json option:

{
  "autoCompact": {
    "customInstructions": "Preserve all file paths, function names, error messages, debugging steps, and architectural decisions. Do not abstract or generalize."
  }
}

1. CLAUDE.md section that gets respected during auto-compact:

# Auto-Compact Instructions
When compacting, preserve:
- All file modifications with exact paths
- Error messages verbatim
- Debugging steps taken
- Code patterns and architectural decisions

1. PreCompact hook enhancement: Allow custom_instructions to be injected programmatically for auto-triggered compaction (currently only available for manual /compact).

Additional Context

• Related issues: #6689 (--no-auto-compact flag), #11819 (configurable threshold), #3349 (pre/post compact hooks)
• Current PreCompact hook receives empty custom_instructions for auto-trigger, making it impossible to influence summary quality programmatically
• Many power users disable auto-compact entirely due to this limitation, losing the convenience feature

Expected Behavior

Users should be able to define what details matter most for their workflow, and auto-compact should respect these preferences when generating summaries.