/insights command only analyzes 3 sessions out of 2800+, missing most project data

Bug Description

The /insights command generates a report that claims to analyze 2813 sessions, but the actual facet data only contains 3 sessions. This results in a completely inaccurate and misleading report.

Evidence

What the report claims:

• 2813 sessions
• 23,694 messages
• 3217 hours
• 533 commits

What actually exists:

• 17 project directories in ~/.claude/projects/
• 199MB of session data total
• One project alone (Dropkit v2) has 50 session files (~100MB)

What was actually analyzed:

• Only 3 facet files in ~/.claude/usage-data/facets/
• All 3 are from the same Remotion video project
• Zero sessions from the Dropkit v2 project (which has 50 sessions and many git commits)

Steps to Reproduce

1. Have multiple projects with many sessions (I have 17 projects)
2. Run /insights command
3. Check ~/.claude/usage-data/facets/ - only 3 JSON files exist
4. The report shows aggregate stats (sessions, hours) that seem correct, but the qualitative analysis (project areas, friction points, suggestions) is based on only 3 sessions

Expected Behavior

The /insights command should analyze ALL sessions across ALL projects, not just a tiny subset.

Actual Behavior

• Aggregate statistics (session count, hours, messages) appear to be calculated from all data
• But the AI-generated analysis (project areas, friction analysis, suggestions) is based on only 3 sessions
• This makes the entire "Impressive Things You Did", "Where Things Go Wrong", and "Suggestions" sections completely unrepresentative

Impact

This is a critical bug because:

1. Users trust the report to reflect their actual usage
2. The suggestions are based on <0.1% of actual sessions
3. Major projects (like my Dropkit v2 macOS app development) are completely invisible in the analysis
4. The report appears authoritative but is fundamentally misleading

Environment

• Claude Code version: 2.1.31
• OS: macOS Darwin 25.2.0
• Total projects: 17
• Total session data: ~199MB
• Sessions actually analyzed: 3

Screenshots

The report shows detailed time-of-day statistics and tool errors that suggest it read some aggregate data, but the qualitative analysis missed 99.9% of my work.

!Screenshot showing parallel sessions and tool errors statistics

Suggested Fix

The facet generation process should iterate through ALL project directories in ~/.claude/projects/ and process ALL .jsonl session files, not just a small subset.