feat: import historical Claude Code sessions

[davidrobertson]

Summary

• import historical Claude Code sessions from ~/.claude/projects, explicit paths, and --all
• keep Claude parsing, source discovery, format dispatch, event identity, and daemon ingestion in separate modules
• add synthetic coverage for visible messages, completed tools, malformed records, sidechains, metadata, idempotence, CLI selection, and discovery
• add CONTRIBUTING.md and update README, architecture, troubleshooting, Claude setup guidance, and every repository AGENTS.md

Validation

• npm run typecheck
• npm test: 385 tests, 0 failures
• npm run smoke
• npm pack --dry-run
• strict TypeScript no-excuse audit
• independent HEAVY review: approved with no findings

Real Claude Code 2.1.228 QA used copies of 11 local project session files in isolated Agent LCM stores. The first import produced 10 sessions and 91 events with 0 rejected records and 0 failures. The second import produced 91 duplicates and no new events. Session listing and Claude-scoped MCP search succeeded. Source hashes were unchanged, both test daemons were stopped, and all copied data and temporary stores were removed.

Scope and risk

The importer reads primary project JSONL files and skips metadata, thinking blocks, sidechains, subagents transcripts, orphan tool results, and incomplete tool calls. It does not read ~/.claude/history.jsonl or rely on Claude's session index.