AI enrichment pipeline with auto-enrich reports

[abpai]

Summary

• Project enrichment pipeline (wildcard enrich): gathers git context, redacts sensitive content, sends to Codex CLI (gpt-5.4-mini) for synthesis, stores structured summaries per project/time window
• Auto-enrich in reports: wildcard report automatically runs enrichment before generating, so reports are always rich without a manual step. Skip with --no-enrich
• Model defaults refresh: updated stale model references (llama3.2 → qwen3.5:latest, o3 → gpt-5.4-mini placeholders)
• SQLite busy_timeout fix: 5s → 30s to prevent lock errors when report and watch daemon write concurrently
• Docs reorg: moved secondary docs under docs/, removed old scratch demos

Key design decisions

• Agent runs in /tmp with --sandbox read-only — no repo filesystem access (trust boundary)
• Redaction layer strips .env, credentials, private keys, .gitignore patterns before agent sees context
• git log --all captures activity across all branches
• Window-scoped storage UNIQUE(project_key, start_day, end_day) — daily and weekly coexist
• Context hash enables idempotency — unchanged git state skips the agent call
• Device-aware: enrichment injection skipped for non-local device filters

Test plan

• bun run check passes (prettier + typecheck)
• bun test — 72 tests pass
• wildcard enrich --dry-run shows qualifying projects
• wildcard enrich produces summaries stored in DB
• wildcard report --today auto-enriches then generates report
• wildcard report --no-enrich skips enrichment
• wildcard report --no-ai skips both AI and enrichment
• Without codex installed: report works normally, enrichment silently skipped