Problem

When Claude Code compacts conversation context, all dynamic state injected by LoadContext at session start is lost:

• Active work sessions and their progress
• Algorithm phase and ISC criteria
• Learning signal trends
• Key behavioral rules that prevent common violations

The model continues operating but has amnesia about what it was doing, leading to:

• Restarting completed work
• Losing Algorithm phase mid-execution
• Violating rules it was following before compaction

Related to #798 (PostCompactRecovery) and #793 (compaction context loss).

Proposed fix

Add a PostCompact hook that re-injects critical state from disk:

1. Active work sessions — from MEMORY/STATE/work.json (last 48h)
2. Learning signal trends — from MEMORY/STATE/learning-cache.sh
3. Key behavioral rules — hardcoded essentials that prevent common post-compaction violations
4. Algorithm state — current phase, ISC criteria, PRD path from most recent MEMORY/WORK/ entry

All data comes from small cached files — runs in <50ms with no inference calls.

Output is a <system-reminder> block written to stdout, which Claude Code injects into the conversation.